Datasets:

InternScience
/

SGI-DeepResearch-Lite

idx stringlengths 21 21	question stringlengths 142 2.31k	steps listlengths 3 30	answer stringlengths 1 45	discipline stringclasses 10 values	direction stringlengths 15 102	type stringclasses 4 values
SGI_DeepResearch_0000	In GW150914's time–frequency ridge of the waveform spectrogram, three points are given ((t1, f1)=(-0.19118234 s, 35 Hz)，(t2, f2)=(-0.04541778 s, 60 Hz)，(t3, f3)=(-0.01540456 s, 90 Hz)). Based on the data you got, with t_c = 0, assuming an equal mass ratio (η = 0.25), compute the total mass M of the final black hole mass. Provide the total mass M in solar masses (M⊙), rounded to one decimal place, and show the calculation process	[ "From the time–frequency ridge in the spectrogram, read three (t, f) points.\",", "Transform each frequency to X = f^{-8/3} and each time to Y = -t.\",", "Perform a least-squares fit of Y versus X to determine the slope K.", "Relate the slope to the chirp mass using K = (5/256) (G M_c / c^3)^{-5/3} π^{-8/3} a...	62.0	astronomy	gravitational_wave_detection_and_parameter_estimation	properties
SGI_DeepResearch_0005	In the analysis for its first catalog (Catalog 1), the Canadian Hydrogen Intensity Mapping Experiment Fast Radio Burst Project (CHIME/FRB) estimated the all-sky event rate of Fast Radio Bursts (FRBs). This estimate applies to FRBs with a fluence greater than 5 Jy ms, a dispersion measure (DM) above 100 pc cm^{-3}, and a scattering timescale at 600 MHz of less than 10 ms. Subsequently, the team published an Erratum correcting two systematic errors in the original calculation. It was noted that the corrected all-sky rate is lower than the value originally published. Based on this information, calculate the central value of the all-sky event rate originally published by the CHIME/FRB team, prior to the erratum. The result should be in units of bursts sky^{-1} day^{-1}, rounded to the nearest integer.	[ "Search for the relevant paper 'The First CHIME/FRB Fast Radio Burst Catalog' and Erratum", "Recognize the corrected event rate center value is 525 bursts sky^-1 day^-1", "Recognize the dispersion 35%", "Calculate origin event rate is 525 \\times (1-35.9%) \\approx 820" ]	820	astronomy	fast_radio_burst_detection_and_localization	properties
SGI_DeepResearch_0010	One core of the ZTF sky survey strategy is to maximize sky coverage. For this purpose, ZTF has designed an observation scheme consisting of the primary grid and the secondary grid. Please search for relevant papers and understand the survey strategy of ZTF, and answer: When ZTF completes the observations of the primary grid and the secondary grid, what is the final combined fill factor it achieves within the survey footprint? Please provide the result in the form of a percentage, retaining one decimal place.	[ "Search for relevant paper The Zwicky Transient Facility: System Overview, Performance, and First Results", "Understanding the core of the problem: The problem requires the final sky survey filling factor after combining the primary and secondary grids. This is a final indicator of the effectiveness of the observ...	99.2	astronomy	real-time_optical_transient_survey_based_on_ztf	micro-experiments
SGI_DeepResearch_0014	Monolayer molybdenum disulfide (MoS₂), a paradigmatic two-dimensional transition metal dichalcogenide, exhibits complex lattice dynamics governed by electron-phonon coupling and substrate interactions. In the context of 2D materials like monolayer MoS₂, lattice vibrations (phonons) can be modeled as damped harmonic oscillators due to substrate interactions. Using the Class SR-derived damped harmonic oscillator equation \[ x(t) = A e^{-kt} \cos(\omega t) \], where \( A = 1.0 \), \( k = 0.2 \, \text{s}^{-1} \), and \( \omega = 1.389 \, \text{rad/s} \) are parameters calibrated via DFT+U for MoS₂/substrate systems, what is the displacement \( x(t) \) at \( t = 1.5 \, \text{s} \) after accounting for electron-phonon coupling corrections? The coupling introduces a phase shift \( \Delta\phi = \pi/6 \) and amplitude scaling factor \( \lambda = 0.95 \). Provide the value with 3 decimal places.	[ "Find paper Class Symbolic Regression: Gotta Fit ’Em All", "Substitute parameters: \\( A = 1.0 \\), \\( k = 0.2 \\, \\text{s}^{-1} \\), \\( \\omega = 1.389 \\, \\text{rad/s} \\), \\( t = 1.5 \\, \\text{s} \\).", "Compute exponential decay: \\( e^{-0.3} \\approx 0.7408 \\).", "Calculate cosine term without pha...	0.128	astronomy	formula_regression	properties
SGI_DeepResearch_0018	In the dictionary learning module of the CASTER framework, each frequent substructure \(C_i\) is mapped to a basis vector \(b_i\) by an encoder, forming a dictionary matrix \(B \in \mathbb{R}^{50 \times 1722}\). For the drug combination of sildenafil and isosorbide mononitrate, the projection coefficient \(r \in \mathbb{R}^{1722}\) of the latent representation \(z \in \mathbb{R}^{50}\) on the dictionary B is obtained by an optimization method.It is known that the cosine similarity between the basis vector \(b_{\text{nitrate}}\) of the nitrate group (\(O = N^+\)) and z is 0.85, the \(L_2\) norm of \(b_{\text{nitrate}}\) is 1.2, and the \(L_2\) norm of z is 0.8. If the average cosine similarity of all basis vector pairs in B is 0.15, and the column vectors of B are approximately orthogonal, what is the approximate relative importance weight of the nitrate group in the projection coefficient r (i.e., \(r_{\text{nitrate}}\))? (Regularization parameter \(\lambda = 1 \times 10^{-5}\); keep two decimal places for the answer.)	[ "1.Find paper 'CASTER: Predicting Drug Interactions with Chemical Substructure Representation'.", "2.Recall the Ridge Regression Formula,", "According to Page 4, right column, Equation (7) in the paper, the projection coefficients are computed as: r* = (BᵀB + λI)⁻¹Bᵀz where B is the dictionary matrix of basis v...	0.48	chemistry	molecular_interaction	properties
SGI_DeepResearch_0027	In computational chemistry, the accurate parsing of a molecule’s structure is fundamental to predicting its properties. A critical structural attribute is aromaticity, and its determination often follows Hückel’s rule. Consider the neutral molecule, an isomer of Naphthalene, represented by the following SMILES string: c1cccc2cccc-2cc1 For the entire conjugated system of this molecule to be considered aromatic, how many π-electrons in total must its π-electron system contain? Provide the answer as a single integer.	[ "Find the article title \"DrugAgent: Automating AI-aided Drug Discovery Programming through LLM Multi-Agent Collaboration\"", "Parse the SMILES Structure:", "The SMILES string c1cccc2cccc-2cc1 describes the molecule Azulene, a bicyclic conjugated system formed by the fusion of a five-membered ring and a seven-m...	10	chemistry	molecular_property_prediction	properties
SGI_DeepResearch_0032	According to the Gutenberg–Richter law, the frequency–magnitude distribution should follow a straight line on a log–linear plot. However, when comparing the PAL catalog obtained through the PALM monitoring framework with the SCSN catalog shows a deviation in the SCSN catalog. Events smaller than which magnitude are considered incomplete in the SCSN catalog? Please correct the last answer to one decimal places.	[ "Find paper 'An Earthquake Detection and Location Architecture for Continuous Seismograms: Phase Picking, Association, Location, and Matched Filter (PALM)'.", "Navigate to the 'Detection results' section on page 9.", "Locate the sentence that describing the frequency–magnitude comparison.", "Get the sentence ...	3.4	earth	seismic_wave_detection	data
SGI_DeepResearch_0036	Accurate earthquake location relies on correctly associating seismic phases. LOC FLOW adopt the REAL algorithm to associate machine-learning-based P and S picks into specific earthquakes due to its flexible parameters and computational efficiency. After the phase association process, which method is used by LOC FLOW to preliminarily relocate the associated earthquakes and improve the initial location precision? Provide the method name directly.	[ "Find paper 'LOC-FLOW: An End-to-End Machine Learning-Based High-Precision Earthquake Location Workflow'.", "Navigate to the page 4.", "Locate the part in the text that describes phase association.", "Find the sentence that describes using the simulated annealing method to preliminarily relocate earthquakes a...	simulated annealing method	earth	seismic_wave_detection	macro-experiments
SGI_DeepResearch_0040	Based on the analysis of global Ocean Heat Content (OHC) change in the upper 2000 meters using the IAP/CAS dataset, by what percentage did the average annual increase rate during the 1986–2021 period increase compared to the average annual increase rate during the 1958–2021 period? Please round the result to two decimal places and use the percentage sign (%) as the unit.	[ "Find paper 'Another Record: Ocean Warming Continues through 2021 despite La Niña Conditions'.", "Determine the long-term warming rate (R1): Identify the mean annual increase rate of global upper 2000m OHC for the 1958–2021 period using the IAP/CAS dataset.", "R1=5.7 ZJ yr−1.", "Determine the recent warming r...	59.65	earth	ocean_heat_content	macro-experiments
SGI_DeepResearch_0044	For the atmospheric state encoding scheme adopted for medium-range forecasting tasks within the field of atmospheric science, where the most common model inputs include surface layers and upper-air layers (typically 13 pressure levels), if this scheme utilizes the Geopotential (Z) variable across multiple pressure layers along with other surface and upper-air fields to define its dimension, what is the percentage of the total channel dimension allocated to the Geopotential variable, relative to the total channel dimension of all atmospheric variables? Please express the final result to two decimal places, using the percentage sign (%).	[ "Find paper \"Generalizing Weather Forecast to Fine-grained Temporal Scales via Physics-AI Hybrid Modeling\".", "Find the number of pressure levels (P) used to represent the upper-air atmospheric variables.", "According to Table 2, the model uses 13 pressure levels (e.g., 50, 100, ..., 1000 hPa).", "Calculate...	18.84	earth	atmospheric_differential_equation	properties
SGI_DeepResearch_0049	In a meteorological model for estimating tropical cyclone (typhoon) wind pressure, three satellite channels (IR, WV, and PMW) report wind speed estimation errors in knots (kts): 9, 10, and 13, respectively. These errors are used to derive a “Pressure Anomaly Index” (PAI) to quantify the potential distortion in wind pressure calculations, where wind pressure is proportional to the square of wind speed. The PAI is computed through the following layered process: Base Pressure Distortion (BPD): The quadratic mean (RMS) of the three wind speed errors. Channel Harmony Ratio (CHR): The product of the three errors divided by their greatest common divisor (GCD). If the GCD is 1, double the product before proceeding. Anomaly Threshold (AT): The sum of the unique prime factors (counting multiplicity) of the median wind speed error. Conditional Adjustment: Compare the integer part of BPD to AT. If floor(BPD) > AT, then PAI = (CHR / AT)^2 rounded to the nearest integer. Otherwise, PAI = the number of digits in CHR minus floor(BPD / AT). What is the Pressure Anomaly Index (PAI)?	[ "1.Find the article title \"Tropical Cyclone Intensity Estimation Using Multidimensional Convolutional Neural Network From Multichannel Satellite Imagery\"", "2. Calculate Base Pressure Distortion (BPD)", "Action: Compute the quadratic mean (RMS) of the three wind speed errors (9, 10, 13). Formula: sqrt( (9² + ...	111747	earth	typhoon_wind_pressure_relationship	macro-experiments
SGI_DeepResearch_0053	A typhoon intensity probability estimation system integrates two satellite observation sources: 1. Infrared (IR): The estimated value is \(V_I \sim N(\mu_I, \sigma_I^2)\), with a base variance of \(\sigma_{I0}^2 = 4^2\) kt². 2. Microwave (MW): The estimated value is \(V_M \sim N(\mu_M, \sigma_M^2)\), with a base variance of \(\sigma_{M0}^2 = 6^2\) kt². The system uses a conditional weight fusion method to calculate the final 'optimal fused estimate', which is a weighted average of the source estimates. The weighting and fusion rules are as follows: 1. Calculate Initial Weights: The initial weights are calculated based on the base variances: \(w_I^{(0)} = \frac{\sigma_{M0}^2}{\sigma_{I0}^2 + \sigma_{M0}^2}\) and \(w_M^{(0)} = 1 - w_I^{(0)}\). 2. Variance Amplification Rule: First, check the difference between the two source estimates, \(\|\mu_I - \mu_M\|\). - If the difference is greater than 10 kt, a penalty mechanism is activated: An initial fused value \(V_{\text{init}} = w_I^{(0)} \cdot \mu_I + w_M^{(0)} \cdot \mu_M\) is calculated. The source (\(\mu_I\) or \(\mu_M\)) that is farther from \(V_{\text{init}}\) is identified as the 'deviant source', and its variance is temporarily amplified by a factor of 1.5 (used only for this weight recalculation). 3. Final Fusion: The weights \(w_I', w_M'\) are recalculated using the adjusted (or original, if no adjustment was made) variances. The final weighted average fusion is then performed: \(\mu_{\text{fusion}} = w_I' \cdot \mu_I + w_M' \cdot \mu_M\). In a specific observation, the system gets \(\mu_I = 82\) kt and \(\mu_M = 95\) kt. According to this rule, what is the optimal fused estimate for this observation (in kt)? (Round the result to the nearest integer.)	[ "1. Find the article: 'Probabilistic Estimation of Tropical Cyclone Intensity Based on Multi-Source Satellite Remote Sensing Images'.", "2. Calculate the initial weights. This step corresponds to the concept of inverse-variance weighting, a fundamental principle in data fusion discussed within Section 2.3 'Featur...	84	earth	typhoon_wind_pressure_relationship	data
SGI_DeepResearch_0057	In the semi-empirical NDVI model of the Relative Abundance (RA) algorithm, for a temperate shrub vegetation sample area, the annual maximum NDVI (NDVIₚ,max) values of 10 pure shrub pixels in this area (sorted in ascending order) are: 0.82, 0.85, 0.86, 0.88, 0.90, 0.91, 0.92, 0.93, 0.94, 0.95 (unit: dimensionless). The NDVI∞ (NDVI for full vegetation cover) of shrubs is the 90th percentile of NDVIₚ,max. Meanwhile, it is known that the NDVI value of a remote sensing pixel in this area is 0.75 (dimensionless), the bare soil NDVI (NDVIₛ) is 0.05 (dimensionless, Zeng et al., 2000), and the ratio of the canopy gap fraction attenuation coefficient to the NDVI extinction coefficient (Kₚ/Kᵥᵢ) is 0.6175 (dimensionless, Baret et al., 1995). Calculate the green vegetation coverage (fₙ) of this pixel using the semi-empirical NDVI model formula \(f_{c}=1-\left(\frac{NDVI - NDVI_{\infty}}{NDVI_{s} - NDVI_{\infty}}\right)^{K_{p}/K_{VI}}\). Requirements: NDVI∞ shall be rounded to two decimal places during the calculation, and the final fₙ result shall be rounded to three decimal places (dimensionless).	[ "Locate the paper titled \"Remote sensing algorithms for estimation of fractional vegetation cover using pure vegetation index values: A review\";", "Navigate to the \"2 Background and issues\" chapter, find the \"2.1 Semi-empirical NDVI model\" subsection, and confirm the formula of the semi-empirical NDVI model...	0.592	earth	vegetation_coverage_rate	micro-experiments
SGI_DeepResearch_0063	A certain glacier had a volume of \(V_0\) \(\text{km}^3\) and a surface area of \(S_0\) \(\text{km}^2\) in 2000, satisfying the volume-area relationship:\(V = 0.04 \cdot S^{1.5}\)During 2000–2020, the average mass balance of this glacier was \(B = -0.6\) \(\text{m w.e. a}^{-1}\).Assume that the surface area remains unchanged during this period, then the volume in 2020 can be calculated as \(V_1\) according to the mass balance.However, the measured volume in 2020 is \(1.028 V_1\).If the difference between the measured value and the predicted value is entirely caused by the linear change of the surface area, find the average annual surface area change rate of this glacier during 2000–2020 (\(%/\text{a}\)).(The ice density is taken as 917 \(\text{kg m}^{-3}\), and the answer is rounded to two decimal places.)	[ "1.Find paper 'Global glacier volume projections under high-end climate change scenarios'.", "2.Calculate the volume change caused by mass balance (surface area unchanged)\\(\\Delta V_{\\text{mb}} = \\frac{B \\cdot S_0 \\cdot 20}{917}\\)Basis: The conversion relationship between mass balance and volume change in ...	-0.26	earth	glacier_estimation	micro-experiments
SGI_DeepResearch_0067	Based on the study by Wang et al. (2010) in the Beijing suburbs, the ozone production efficiency (OPE) was 6.5 when NOz < 10 ppbv, and decreased to 2.7 when NOz ≥ 10 ppbv. If an air parcel had an Ox (O₃ + NO₂) increase of 52 ppbv at NOz = 8 ppbv, what would the Ox increase be when NOz rises to 15 ppbv? Round your answer to the nearest whole number. (Unit: ppbv)	[ "1.Find the article title \"Ozone pollution in China: A review of concentrations, meteorological influences, chemical precursors, and effects\"", "2.Calculate the initial amount of oxidized nitrogen oxides (based on data in Table 2, Page 8):", "Using initial OPE₁ = 6.5 and ΔOx₁ = 52 ppbv,", "ΔNOz = ΔOx₁ / OPE...	22	earth	ozone_pollution_and_its_causes	micro-experiments
SGI_DeepResearch_0072	In atmospheric chemistry, surface ozone concentration is a key indicator for assessing air quality, and its changes reflect the effectiveness of pollution control policies. Based on observational data from the China National Environmental Monitoring Center for summers from 2013 to 2017, what is the percentage increase in observed surface MDA8 O3 concentration from 2013 to 2017? Please provide the answer as a percentage with one decimal place.	[ "Locate the paper 'Worsening urban ozone pollution in China from 2013 to 2017 – Part 1", "Navigate to Table 2 on page 6: Evaluation results for the air pollutants in China", "Extract the observed mean (OBS) MDA8 O3 for 2013 as 50.9 ppbv", "Extract the observed mean (OBS) MDA8 O3 for 2017 as 56.3 ppbv", "Cal...	10.6	earth	ozone_pollution_and_its_causes	macro-experiments
SGI_DeepResearch_0076	A dual-satellite remote sensing system performs collaborative observation over an urban area. The Optical Satellite (Sat-O) carries a push-broom sensor with a focal length f = 8 m, pixel size p = 6 μm, side-view angle θ = 28°, and orbital altitude H_orbit = 600 km. The SAR Satellite (Sat-S) operates in X-band (wavelength λ = 0.031 m), with an incidence angle φ = 42°, range bandwidth B_r = 200 MHz, and employs multi-look processing (azimuth looks L_a = 3, range looks L_r = 2) to reduce speckle noise. A flat-roof building is monitored, with a measured height of H_b = 75 m. On the optical image acquired by Sat-O, the measured roof projection shift is Δ = 26 pixels. From the SAR image acquired by Sat-S, the building wall baseline is extracted by detecting the double-bounce scattering bright line. To correct the projection shift in the optical image via image registration, the registration accuracy is required to be within one resolution cell of the SAR image. Calculate: In the range direction, what is the minimum integer number of pixels N_min by which the roof edge in the optical image must be shifted?	[ "1.Find paper 'A method of correction building roofs offset using wall baselines fromSAR imagery'.", "2.Introduction (Page 1)", "Calculate the Ground Sampling Distance (GSD) of the optical image", "GSD = (H_orbit / f) × p = (600 × 10^3 / 8) × (6 × 10^(-6)) = 75000 × 6 × 10^(-6) = 0.450 m/pixel", "3.Introduc...	89	earth	emission_inversion_based_on_satellite_remote_sensing_and_four-dimensional_variational_method	micro-experiments
SGI_DeepResearch_0080	Using a chemical transport model and satellite NO₂ column concentration to retrieve surface NOₓ emissions, the mass conservation equation for a certain grid is simplified as [NO₂]ₛₐₜ = E/k + [NO₂]ᵦₖ₉. Where: [NO₂]ₛₐₜ: 1.5 × 10¹⁶ molec/cm² (satellite observation value) E: The emission flux to be determined (molec/cm²/s) [NO₂]ᵦₖ₉: 3.0 × 10¹⁵ molec/cm² (background concentration) Eₚᵣᵢₒᵣ: 2.5 × 10¹¹ molec/cm²/s (a priori emission) [NO₂]ₘₒ𝒹ₑₗ: 1.2 × 10¹⁶ molec/cm² (model simulated concentration when using Eₚᵣᵢₒᵣ) σ_E: 50% (uncertainty of a priori emission) σₛₐₜ: 40% (uncertainty of satellite observation) σ_k: 30% (relative uncertainty of removal coefficient k) Questions: Calculate the value of the removal coefficient k.（Keep the answer to three decimal places） Calculate the "top-down" emission estimate Eₜₒₚ₋ₙₒᵥₙ.（Keep the answer to three decimal places） Calculate the uncertainty of "top-down" emission σₜₒₚ₋ₙₒᵥₙ.（Keep the answer to three decimal places） Calculate the optimal posterior emission Eₚₒₛₜₑᵣᵢₒᵣ.（Keep the answer to one decimal places） Separate the answers with commas	[ "1.Find paper 'Assimilated inversion of NOx emissions over east Asia using OMI NO2 column measurements'.", "2. Calculate the removal coefficient κ", "Formula: κ = Eₚᵣᵢₒᵣ / ([NO₂]ₘₒ𝒹ₑₗ - [NO₂]ᵦₖ₉)", "Calculation:", "[NO₂]ₘₒ𝒹ₑₗ - [NO₂]ᵦₖ₉ = 1.2×10¹⁶ - 3.0×10¹⁵ = 9.0×10¹⁵ molec/cm²", "κ = (2.5×10¹¹) / (9.0...	2.778×10⁻⁵,3.334×10¹¹ ,1.667×10¹¹,2.8×10¹¹	earth	emission_inversion_based_on_local_mass_conservation	micro-experiments
SGI_DeepResearch_0085	Provide the exact technical term that matches each definition below. Answer format: three keywords in lowercase, hyphenate multi-word terms with a single hyphen, use ENGLISH commas, no spaces, no units. Example: term-a,term-b,term-c (1) A voltage stability indicator bounded in [0,1], where smaller values denote a more stable system; computed using YBUS partitions (YLL and YLG) and defined per load bus, with the objective often to minimize its maximum value across all load buses. (2) The sum, over all load buses, of the absolute deviation of each bus voltage magnitude from 1.0 p.u. (3) The probability density function used to model wind speed uncertainty in the study, characterized by a single scale parameter σ.	[ "Match definition (1) to the well-known stability metric whose value lies between 0 and 1 and decreases as stability improves; it is constructed from YLL and YLG derived from the YBUS matrix and commonly optimized via its maximum across PQ buses → l-index.", "Match definition (2) to the index defined as the sum o...	l-index,voltage-deviation,rayleigh	energy	optimal_power_flow_calculation	micro-experiments
SGI_DeepResearch_0090	Wavelet Packet Decomposition (WPD) is a signal processing technique used to break down a signal into various frequency components. A crucial step in this process is the selection of a 'mother wavelet' function, which acts as the prototype for all the wavelet basis functions.\n\nThe choice of mother wavelet influences the decomposition's characteristics. The methodology section of a study using WPD must specify the chosen function, such as those from the Daubechies, Coiflets, or Symlets families.\n\nA WPD-LSTM forecasting model is proposed where the PV power series is first decomposed using a function from the Daubechies wavelet family in 《A hybrid deep learning model for short-term PV power forecasting》.\n\nWhat specific type and order of the Daubechies-type wavelet function was utilized as the mother wavelet for the decomposition?\n\nProvide the answer as a single abbreviated string.,	[ "Find paper 'A hybrid deep learning model for short-term PV power forecasting',", "Navigate to Section 2.4, 'WPD-LSTM forecasting model'.,", "Read the first paragraph of the section to find the description of the wavelet function.,", "Locate the sentence: 'A frequently-used Daubechies-type wavelet function of...	db3	energy	fengguang_new_energy_power_forecasting	micro-experiments
SGI_DeepResearch_0098	In the CraisglistBargain negotiation dataset, what is the average percentage increase in Sale-to-List Ratio when PPDPP is compared to the ProCoT baseline method? Express your answer as a percentage with two decimal places, using % as the unit.	[ "Find paper Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents, from Table 3 in the paper, identify the Sale-to-List Ratio (SL%) for PPDPP on CraisglistBargain: 0.3376", "From Table 3, identify the Sale-to-List Ratio for ProCoT baseline on CraisglistBargain: 0.2486", "Calculate the ab...	35.80	information	dialogue_system	properties
SGI_DeepResearch_0105	Existing code benchmarks (e.g., HUMANEVAL) heavily rely on manually constructed test-cases to evaluate LLM solutions. However, these tests often fall short in capturing all possible scenarios. As such, EvalPlus transforms HUMANEVAL to HUMANEVAL+ byadding 80× unique test-cases and fixing incorrect ground-truth solutions in HUMANEVAL. In HUMANEVAL+, 83 out of the 164 programming tasks are annotated with hand-crafted contracts. And the pass@k results when evaluating the StableLM model using both HUMANEVAL and HUMANEVAL+.When the sample size is 10 and 100, the StableLM model achieves a better pass rate on the original HUMANEVAL benchmark than on the more stringent HUMANEVAL+ benchmark. So, when the sample size k is 1 and the temperature T is set to 0.2, does the pass@1* performance of the StableLM model decrease when switching from the HUMANEVAL to the HUMANEVAL+ benchmark?(Please answer with one word, yes or no).	[ "Find 'Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation'.", "Locate section 3 'Evaluation'.", "Find 'Evaluation of LLMs'. Introducing the pass@k when evaluating LLMs using both the base HUMANEVAL and HUMANEVAL+.", "Observe Table 3 and locate th...	no	information	code_generation	micro-experiments
SGI_DeepResearch_0110	In the field of phaseless spherical near-field measurement for antenna testing, the dual-scan technique retrieves the radiation pattern of the Antenna Under Test (AUT) using near-field amplitude data from two spherical surfaces with different radii, and it needs to be combined with Spherical Wave Expansion (SWE), Rayleigh far-field distance, and geometric parameter analysis of the AUT. Among them, the 'truncation index N' specifically refers to the ceiling value calculated based on the wavenumber k (k=2π/λ, where λ is the operating wavelength) and the minimum sphere radius r₀ enclosing the AUT; the 'minimum sphere radius r₀' is defined as the radius of the smallest sphere that completely encloses the AUT, which needs to be calculated based on the enclosing box dimensions of the AUT; the formula for the 'Rayleigh far-field distance r_far' is r_far=2×(2r₀)²/λ. The formula for the truncation index is N=⌈kr₀⌉+10; 2. When the mmVAST antenna operates at 37.8 GHz, its enclosing box dimensions are 530mm×230mm×440mm, the truncation index N=250, and the Rayleigh far-field distance r_far=91m. First, calculate the theoretical maximum value of r₀, then determine the unique value of kr₀, and further derive the operating wavelength λ of the antenna. The result is required to retain 6 decimal places, unit: meters.	[ "Locate the paper 'Numerical and Experimental Investigation of Phaseless Spherical Near-Field Antenna Measurements';", "Find Section IV.B that describes the mmVAST antenna, and extract key data: enclosing box dimensions 530mm×230mm×440mm, truncation index N=250, Rayleigh far-field distance r_far=91m;", "Convert...	0.007796	information	sensor_spatial_characteristics_phase-free_reconstruction	micro-experiments
SGI_DeepResearch_0115	In the validation cohort, the durable clinical benefit (DCB) rate was 83% in the high mutation burden group (n=9) and 22% in the low burden group (n=9). A patient originally classified in the low burden group was found to have a POLE mutation and was reclassified into the high burden group. If the POLE mutation increases this patient's probability of benefit from 22% to 90%, calculate the difference in overall DCB rates between the reclassified high and low burden groups (high group rate − low group rate), the unit is percentage, the result should be rounded to three decimal places.	[ "Find the paper Mutational landscape determines sensitivity to PD-1 blockade in non–small cell lung cancer", "Extract validation cohort baseline data:", "High burden group: n=9, DCB=83%", "Low burden group: n=9, DCB=22%", "*Corresponding text: Page 1, right column, paragraph 5: \"DCB 83% versus 22%, Fisher’...	56.750	life	tumor_immunotherapy	data
SGI_DeepResearch_0119	A research team plans to use a single computing node to predict the complete structures of 1,500 antibody sequences. If the total computation time must not exceed 10 hours, which prediction method should be selected? Verify through calculation and provide the name of the method.	[ "1.Find the article title \"Fast, accurate antibody structure prediction from deep learning on massive set of natural antibodies\"", "2.Locate the runtime comparison section in the \"Minimal refinement yields faster predictions\" part of the paper (Page 7) and extract the average runtime per sequence for each met...	IgFold	life	ai-assisted_antibody_design	properties
SGI_DeepResearch_0123	In the EF-hand calcium-binding motif grafting experiment, from which PDB entry was the minimal EF-hand sequence used for Tb³⁺ binding derived?	[ "1.Fing the article title \"De novo design of immunoglobulin-like domains\"", "2.(Page 9, Left Column, Second Paragraph) Locate the \"Methods\" section describing the design of EF-hand calcium-binding motifs.", "3.(Page 9, Left Column, Second Paragraph) Identify the sentence stating: \"A minimal EF-hand motif f...	1NKF	life	ai-assisted_antibody_design	data
SGI_DeepResearch_0128	A spherical protein consists of 180 amino acids, with its hydrophobic core accounting for approximately 40% of the total volume. Given that the average volume per amino acid is 110 Å³ and the packing density of the protein is 0.75, if the apparent volume of an amino acid in the hydrophobic core is 140 Å³, approximately how many amino acids are contained in the hydrophobic core? (Round to the nearest integer)	[ "1.Find the article title \"Deep Learning-Based Advances in Protein Structure Prediction\"", "2.Calculate total protein volume using packing density,", "calculation: V_total = (180 × 110) / 0.75 = 26400 Å³,", "rationale: Standard volume calculation using atomic volume and packing density", "3.Calculate hydr...	75	life	protein_structure_prediction	properties
SGI_DeepResearch_0134	PrismNN identified the metrics for the level 6 high-risk group in risk stratification. Based on this indicator, what is the estimated baseline (expected) three-year incidence rate for the level 6 high-risk group (expressed as a percentage, rounded to two decimal places)?	[ "Find the paper \"A pancreatic cancer risk prediction model (Prism) developed and validated on large-scale US clinical data.\"", "Read the original table values for level 6:", "PrismNN level 6: PPV (TrxPop. Est.) = 8.62%.", "PrismNN level 6: SIR (TrxPop. Est.) = 96.0.", "Variable transformation and algebrai...	0.09	life	early_screening_and_risk_stratification_of_pancreatic_cancer	macro-experiments
SGI_DeepResearch_0138	In the TGNN module of the HIGH-PPI model, for a PPI network containing m proteins, its adjacency matrix \(A_c\) is known. A 3-layer GIN (Graph Isomorphism Network) is used for feature propagation, where the MLP of each GIN layer is a single-layer linear transformation (without bias) and the output dimension remains d. When training stabilizes, it is known that a protein node v and its neighbor set \(N(v)\) satisfy the following condition: the L2 norm of the difference between the final embedding \(h_v^{(3)}\) of node v and the mean of the initial features of its first-order neighbors (i.e., \(h_v^{(3)} - \text{mean}(\{h_u^{(0)} \mid u \in N(v)\})\)) is constrained to be less than or equal to \(\delta\). Given that the degree of node v is k, and all initial feature vectors \(h_u^{(0)} \in \mathbb{R}^d\) are mutually orthogonal with a modulus (norm) of 1, find the upper bound of the maximum singular value of the weight matrix \(W^{(l)}\) of each MLP layer in GIN (expressed in terms of \(\delta\), k, and d).	[ "1.Find the article title \"Hierarchical graph learning for protein–protein interaction\"", "2.Define the GIN Propagation Rule (Based on Page 9, \"TGNN for learning PPI network information\")", "The propagation rule for a node v at layer l in a GIN is given by:", "h_v^(l) = MLP^(l)( (1+ε) · h_v^(l-1) + Σ_{u∈N...	δ^(1/3) / k	life	protein-protein_interaction_prediction	properties
SGI_DeepResearch_0144	Adoption of high-content omic technologies in clinical studies, coupled with computational methods, has yielded an abundance of candidate biomarkers. However, translating such findings into bona fide clinical biomarkers remains challenging. The researchers introduce a general machine learning method that identifies a sparse, reliable set of biomarkers by integrating noise injection and a data-driven signal-tonoise threshold into multivariable predictive modeling, which enable biological interpretation of complex predictive models in multi-omic integration tasks. Which technology's accuracy does the effectiveness of this method depend on? Please answer with a phrase.	[ "Find paper Discovery of sparse, reliable omic biomarkers with Stabl.", "Check the sentence:a general machine learning method that identifies a sparse, reliable set of biomarkers by integrating noise injection and a data-driven signal-tonoise threshold into multivariable predictive modeling.", "Define the/this ...	(the) artificial feature generation technique	life	biomarker_discovery	data
SGI_DeepResearch_0151	In the DeepSTARR model, a human enhancer contains two identical p53 core motifs (RRRCWWGYYY) at positions +50 and +150. Experimental data show: Mutating the +50 motif alone reduces H3K27ac signal to 35% of wild-type Mutating the +150 motif alone reduces H3K27ac signal to 82% of wild-type DNase I footprinting shows TF binding at the +50 motif but no binding at the +150 motif Changing the 5' flanking sequence of the +150 motif from "GGG" to "CTC" confers TF binding ability Known effects of flanking sequences on p53 binding: Optimal flank "GGG": increases binding affinity by 8-fold Suboptimal flank "CTC": increases binding affinity by 3-fold Random flank: binding affinity = 1 (baseline) Assume H3K27ac signal strength is proportional to p53 binding affinity, and total signal equals the sum of both motifs' binding affinities. If the +50 motif's flank is changed from "GGG" to "CTC" and the +150 motif's flank is changed from "GGG" to "CTC", what is the predicted H3K27ac signal as a percentage of wild-type?The result retains the integer.	[ "1.Find the article title \"DeepSTARR predicts enhancer activity from DNA sequence and enables the de novo design of synthetic enhancers\"", "2.Determine wild-type binding affinities", "+50 motif: flank \"GGG\" → affinity = 8 (Article: Fig. 4 & related text – flanking sequences significantly influence motif imp...	67	life	regulatory_element_design	micro-experiments
SGI_DeepResearch_0155	In tandem mass spectrometry, a peptide undergoes fragmentation producing a pair of complementary b-ion and y-ion. The observed m/z of the b-ion is 586.32 (charge z=1), and the m/z of the y-ion is 687.45 (charge z=1). During the fragmentation, a neutral water loss (molecular mass 18.02 Da) occurs. What is the precursor mass of the intact peptide in Da,Round the result to two decimal places.? (Hint: Consider the complementary ion relationship and neutral loss.)	[ "1.Find the article title \"ContraNovo: A Contrastive Learning Approach to Enhance De Novo Peptide Sequencing\"", "2.Confirm the Masses of the b-ion and y-ion", "Calculation Basis: In mass spectrometry, when the charge state z=1, the mass-to-charge ratio (m/z) is numerically equal to the mass of the ion. This i...	1255.75	life	de_novo_protein_sequencing	micro-experiments
SGI_DeepResearch_0159	Using A100 GPU, Casanovo V2 without beam search (w/o Beam) has an inference speed of 227 spectra/s, while π-PrimeNovo without precise mass control (w/o PMC) has an inference speed of 714 spectra/s. The text states that when both models use their respective post-prediction decoding strategies (PMC for PrimeNovo, beam search for Casanovo V2), PrimeNovo is over 28 times faster than Casanovo V2.If Casanovo V2 with beam search takes X hours to process the entire nine-species test set (93,750 spectra), calculate the value of X (round to two decimal places).	[ "1.Find the article title \"π-PrimeNovo: an accurate and efficient nonautoregressive deep learning model for de novo peptide sequencing\"", "2.Determine the inference speed of Casanovo V2 with beam search", "Operation: Extract the speed value of Casanovo V2 with beam search from Figure 2d", "Value: 8 spectra/...	3.26	life	de_novo_protein_sequencing	data
SGI_DeepResearch_0169	MedGemma is a series of foundational medical visual-language models jointly developed by Google Research and DeepMind. They are based on the Gemma 3 architecture (with parameter sizes of 4 billion and 27 billion).For out-of-distribution, by how many relative percentage points does this model improve accuracy on the AgentClinic benchmark and electronic health record (EHR) information retrieval and reasoning tasks compared to baseline models, and by how many relative percentage points does the chest X-ray (pneumothorax) binary classification accuracy improve after fine-tuning? Please separate each answer with a semicolon, express as percentages, and round to one decimal place.	[ "Search keyword：out-of-distribution,", "Find the abbreviation for out-of-distribution is OOD,", "Search keywords：AgentClinic benchmark、EHR information retrieval and reasoning、accuracy", "Locate table 11、14 and 13:", "AgentClinic-MIMIC(OOD)Accuracy(%):MedGemma27B=46.0,Gemma3 27B=35.2;", "EHRQA accuracy res...	30.7;11.2;2.2	life	medical_image_understanding	data
SGI_DeepResearch_0175	A researcher is using mass spectrometry to analyze an unknown small molecule metabolite. He uses the SIRIUS software for molecular formula prediction, which for compounds with a mass below 500 Da has a prediction error rate of less than 10%. The exact mass of the compound he measured is 380.1000 Da.\n\nIt is known that when SIRIUS performs its calculations, the number of different molecular formulas (N) possible for a given mass (M) is proportional to the cube of that mass, i.e., N ∝ M³. It is given that a compound with a mass of 300 Da has approximately 1,000 possible molecular formulas.\n\nCalculate the approximate number of candidate molecular formulas for the researcher's unknown metabolite. (Round your answer to the nearest integer).	[ "1.Find the article title \"MSNovelist: de novo structure generation from mass spectra\"", "2.Establish the Proportional Relationship Model", "Calculation: Define the relationship N = k × M³.", "Source in Paper: Page 1, INTRODUCTION. The text states: \"This approach fails in practice because of the combinator...	2032	life	small_molecule_inference	properties
SGI_DeepResearch_0179	Each NeoVax vaccine dose contained approximately a certain number of personalized immunizing peptides (IMPs) per patient. It is known that among all 124 tested IMPs, 69 induced CD4⁺ T cell responses (i.e., the initial positive rate = 69/124). Among the CD4⁺ ASP (class II) responses that were positive at week 16, 68% remained detectable during long-term follow-up at 28–56 months (i.e., durability = 68%). Now, suppose a patient received a vaccine containing 20 IMPs, where each peptide independently triggers a CD4⁺ response with the same probability as the overall positive rate, and all peptide events are independent. Calculate the probability (expressed as a percentage, rounded to four decimal places and including the “%” unit) that this vaccine dose contains at least one CD4⁺-positive IMP that remains detectable for ≥28 months.	[ "Find the paper \"Personal neoantigen vaccines induce persistent memory T cell responses and epitope spreading in patients with melanoma\"", "The probability that a single IMP induces a CD4⁺ response is: pCD4=69/124", "The conditional probability that this response remains detectable during long-term follow-up ...	99.9925	life	tumor_neoantigen_discovery	data
SGI_DeepResearch_0185	A deep learning-based method for RNA 3D structure prediction performs excellently on natural ribosomal RNA (rRNA) but shows significantly lower accuracy on laboratory-designed synthetic riboswitches. The method's performance is strongly correlated with the depth of the Multiple Sequence Alignment (MSA) for the target RNA. It is known that natural rRNAs possess numerous homologous sequences, yielding very deep MSAs, whereas synthetic riboswitches have virtually no homologous sequences, resulting in very shallow or empty MSAs. What is the most direct and decisive factor causing the method's failure on synthetic riboswitches? (Your answer must be a single two-word bioinformatics term.)	[ "1.Find the article title \"trRosettaRNA: automated prediction of RNA 3D structure with transformer network\"", "2.Identify the Algorithm's Core Dependency", "Corresponding Location in Paper: Page 2, Left Column, Paragraph 2.", "Purpose: To confirm that the core input to the algorithm is the Multiple Sequence...	Coevolutionary signals	life	rna_tertiary_structure_prediction	data
SGI_DeepResearch_0190	Explaining the impact of genomic sequence variations remains a core biological challenge, and developing a model that unifies multimodal prediction, long sequence background, and base pair parsing into one framework. This model takes 1 trillion DNA sequences as input and predicts different genomic trajectories of various cell types. Based on the AlphaGenome scoring method for polyadenylation variant effects, consider a gene with two polyadenylation sites: P1 proximal and P2 distal. Under the reference allele, the predicted RNA seq coverage values are P1ref=200 and P2ref=100. Under alternative allele, the predicted coverage values are P1alt=100 and P2alt=200. Assuming that the segmentation method of PAS is only for the proximal and distal sets, calculate the paQTL score for this variant and keep the result as an integer.	[ "Find paper\"AlphaGenome: advancing regulatory variant effect prediction with a unified DNA sequence model\",", "REF ratio: Rref=P2ref/P1ref=100/200=0.5,", "ALT ratio: Ralt=P2alt/P1alt=200/100=2,", "Calculate absolute log2 fold change: use formula \|log2 (Ralt)-log2(Rref)\|,", "log2(Ralt)=log2(2)=1,", "log2...	2	life	genome_function_prediction	properties
SGI_DeepResearch_0196	An AI-native biotech has 24 AI-discovered molecules that completed Phase I trials, with 21 successes. Historical industry average Phase I success is 40–65%. Among the successful molecules, AI-discovered small molecules account for 3/7. Assume small molecules and non-small molecules have the same Phase I success rate, and among non-small molecules, only monoclonal antibodies (mAbs) can fail due to immunogenicity. In this company’s failed molecules, mAbs account for 1/3 of failures. How many failed non-small-molecule non-mAb molecules failed for reasons other than immunogenicity? Round the result to the nearest integer.	[ "1.Find the article title \"How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons\"", "2.calculation: 24 - 21 = 3,", "explanation: Total failed molecules,", "source:Page 2, right column, paragraph 2 — '24 AI-discovered molecules had completed Phase I trials, of which...	1	life	ai_drug_discovery	micro-experiments
SGI_DeepResearch_0203	What is the mass activity ratio between the Ir₁/CN catalyst and the commercial Pt/C catalyst in the formic acid oxidation reaction (FAOR)? (Provide the calculated integer ratio)	[ "1.Find the article title Iridium single-atom catalyst on nitrogen-doped carbon for formic acid oxidation synthesized using a general host–guest strategy", "2.Locate the mass activity value of Ir₁/CN (12.9 A mg⁻¹) in the Abstract/Introduction section (Page 1, right column, 1st paragraph).", "3.Locate the mass a...	19	material	thermal_electrocatalysis	properties
SGI_DeepResearch_0208	Generative models can learn the probability distribution of a training dataset. A dataset was constructed to train the MUFFSION multimodal conditional diffusion model. Under a pressure range of 5–100 bar at 77 K, target structures were generated corresponding to hydrogen working capacities of 5, 15, 25, and 35 g/L. Which target structure achieves the highest relative adsorption amount? Provide the target adsorption numbers.	[ "Find paper “Multi-modal conditioning for metal-organic frameworks generation using 3D modeling techniques”.", "Understand the experimental setup: The conditional model was trained for hydrogen working capacity (WC), determined by a pressure swing between 5 bar and 100 bar at 77 K,", "Identify the key result fr...	15	material	nano_adsorption_materials	properties
SGI_DeepResearch_0212	All-solid-state batteries require advanced cathode designs to realize their potential for high energy density and economic viability. Integrated all-in-one cathodes, which eliminate inactive conductive additives and heterogeneous interfaces, hold promise for substantial energy and stability gains but are hindered by materials lacking sufficient Li+/e− conductivity, mechanical robustness and structural stability. We know the Crystal structure of Li1.3Fe1.2Cl4: The Li and Fe atoms are bonded to six Cl atoms to form edge-sharing octahedra (Fig. 2c); Most of the Fe atoms sit in Wyckoff position 2a, whereas Li atoms and a small amount of Fe atoms occupy Wyckoff position 4f. In the context of the Crystal structure of Li1.3Fe1.2Cl4, what is the space group of Li1.3Fe1.2Cl4?	[ "Find the paper A cost-effective all-in-one halide material for all-solid-state batteries", "Find the determination of the Crystal structure of Li1.3Fe1.2Cl4", "Determine the space group of the Li1.3Fe1.2Cl4," ]	Cmmm space group	material	chloride_solid-state_electrolyte	properties
SGI_DeepResearch_0219	In nickel-iron-based catalysts, the stability at high current densities can be enhanced by introducing nitrate ions. At what voltage (V vs. RHE) does significant structural rearrangement occur in nitrate-containing nickel-iron-based OER catalysts? Keep the result to two decimal places.	[ "Find the paper Oxyanion Engineering Suppressed Iron Segregation in Nickel–Iron Catalysts Toward Stable Water Oxidation", "Locate paragraphs and sentences in the article concerning the restructuring of nickel-iron-based catalysts.", "Identify the cause of the disappearance of characteristic nitrate ion peaks in...	1.55	material	oxygen_evolution_reaction_catalytic_materials	properties
SGI_DeepResearch_0226	The conjugated carbon double bonds of polyester fatty acid chains modified with vegetable oil can generate free radical self oxidation reactions catalyzed by metal desiccants, which contribute to crosslinking and film-forming properties. The gel ability of the resin can be improved by adjusting the ratio of glycerol and pentaerythritol. Due to the high branching ability of polyols, what properties of the resin will be reduced by the addition of pentaerythritol?	[ "Find paper“NMR and GPC Analysis of Alkyd Resins: Influence of Synthesis Method, Vegetable Oil and Polyol Content”,", "Understand the experimental setup: All alkyd resins contain different weight ratios of glycerol to pentaerythritol", "Understanding the working principle of pentaerythritol: Polyols are highly ...	dispersity	material	krf_resin_polymerization_reaction	properties
SGI_DeepResearch_0239	Within the framework of Differential Privacy, a data holder wishes to release the sum of its user credit scores. Before any computation, each user’s raw score is pre-processed by clipping it to the range [-100, 100]. The institution employs the Laplace Mechanism to add noise to this sum query, aiming to satisfy ε-differential privacy with a privacy budget of ε = 2. The noise is drawn from a Laplace distribution, Lap(b), centered at 0 with a scale parameter b. What is the value of the scale parameter b? Round your result to the nearest integer.	[ "Find the article title \"A Statistical Framework for Differential Privacy\".", "Identify the query function f(X) and its sensitivity S(f). The query is the sum of scores. Sensitivity is defined as the maximum L1 difference in the function's output when one entry in the database is changed. For a sum query over d...	100	mathematics	differential_privacy	micro-experiments
SGI_DeepResearch_0243	An optimization algorithm needs to determine an update quantity z by minimizing a one-dimensional function D(z). The algorithm employs Newton’s Method with Backtracking Line Search. Method and Definitions: First, calculate the Newton direction d = -D'(0) / D''(0). Then, determine the step size λ via a backtracking line search. Starting with λ=1, repeatedly multiply λ by β until the first λ that satisfies the Sufficient Decrease Condition is found. The final update quantity is z = λ * d. Core Formulas: Sufficient Decrease Condition: D(z) - D(0) ≤ -σ * z² Objective Function Change: D(z) - D(0) = zw₃ + (1/2)z² + C [ Σ (max(0,b_j(z)))² - Σ (max(0,b_j(0)))² ] State Evolution: b_j(z) = b_j(0) - z * y_j * x_{j3} Derivatives: D'(z) = w₃ + z - 2C * Σ_{j∈I(z)} b_j(z) * y_j * x_{j3} D''(z) = 1 + 2C * Σ_{j∈I(z)} x_{j3}^2 I(z) is the set of all indices j for which b_j(z) > 0. Given Parameters & Initial Conditions (at z=0): Backtracking Parameters: β = 0.5, σ = 0.1 Model Parameters: w₃ = 0.5, C = 5.0 Initial State Vector: b(0) = [0.8, -0.2, 1.2, 0.6] Parameter Vectors: y = [-1, 1, 1, -1], x_3 = [1.0, 1.0, 0.5, 0.2] Calculate the final update quantity z found by the algorithm. Then, multiply your result by -10000 and round to the nearest integer.	[ "Find the article title \"Coordinate Descent Method for Large-scale L2-loss Linear Support Vector Machines\".", "Calculate the initial derivatives and the Newton direction d. First, determine the set I(0) = {j \| b_j(0) > 0} = {1, 3, 4}. Using this, calculate D'(0) = 3.7 and D''(0) = 13.9. The Newton direction is ...	2662	mathematics	coordinate_descent_optimization_algorithm	micro-experiments
SGI_DeepResearch_0247	You are evaluating the data acquisition requirements for a matrix completion project. According to relevant theory, the minimum number of random samples, m, required to successfully recover an n x n, rank-r matrix satisfying certain 'incoherence' conditions has a theoretical lower bound. You are comparing two theoretical bounds: 1. 'Conservative Theoretical Bound': Derived from this paper's core analysis, its sample complexity is m_con ≥ C_con * μ₀ * n^(6/5) * r * log(n). 2. 'Aggressive Theoretical Bound': Derived from more optimal subsequent research cited in the paper, its sample complexity is m_agg ≥ C_agg * μ₀ * n * r * (log(n))². Given the project parameters: - Matrix dimension n = 1,000,000 (10⁶) - Estimated rank r = 50 - Incoherence parameter μ₀ = 2.5 - Theoretical constants C_con = 10, C_agg = 6 - log is the natural logarithm (ln) Question: Calculate how many times larger the sample requirement of the 'Conservative Bound', m_con, is compared to that of the 'Aggressive Bound', m_agg. Round your result to two decimal places.	[ "Find the article title \"Exact Matrix Completion via Convex Optimization\".", "The core of this problem is to compare two theoretical bounds for the required number of samples in matrix completion. The first bound comes from the paper's core Theorem 2.3, while the second is from subsequent research discussed in ...	1.91	mathematics	matrix_completion	micro-experiments
SGI_DeepResearch_0251	Consider a subset of the Netflix Prize dataset, where the matrix dimension is 471,268 × 1,760, the missing rate is 93.6%, and the true rank is 8. Based on the theoretical convergence guarantee of Corollary 3 under the uniform sampling model in the paper, the following technical assumptions are made: Technical Assumptions: Assumption 1: The matrix A to be recovered has rank k and can be decomposed into A = U∗S∗ᵀ, where U∗ ∈ ℝⁿˣᵏ, S∗ ∈ ℝᵐˣᵏ, ‖S∗‖₂ = 1, and both U∗ and S∗ satisfy the (α, l)-submatrix full rank property, where l/m is the sampling rate, and α > 0. The so-called (α, p)-submatrix full rank property means that for a matrix S ∈ ℝᵐˣᵏ (m > k), each p × k submatrix is column-full rank and the minimum singular value is at least α. Assumption 2: The set of matrices S is restricted to matrices satisfying the (α, l)-submatrix full rank property, where l/m is the sampling rate and α > 0. The problem is considered under the conditions that the regularization parameter 1/γ = 0 and B = Iₘ (identity matrix). Under these assumptions, when the constant C takes the minimum sufficient value of 1, according to the sampling rate r₀ > Cnk log(n)/(mn), what is the minimum number of observations (the number of known entries \|Ω\|) required to ensure that the algorithm converges to the δ-neighborhood of the global optimum with a probability of at least 1 − O(1/n)? The calculation uses log(471,268) ≈ 13.06. Please express your answers in integer form, without using scientific notation.	[ "From paper 'Fast Exact Matrix Completion: A Unified Optimization Framework for Matrix Completion', Step 1: Understanding the Sampling Rate Condition of Corollary 3", "According to Corollary 3, under a uniform sampling model, the sampling rate must satisfy:", "r₀ > Cnk log(n)/(mn)", "where C is a sufficiently...	49240090	mathematics	matrix_completion	properties
SGI_DeepResearch_0255	A 2D nonlinear Burgers' equation, ∂v/∂t = v_xx + v_yy - vv_x - vv_y, is being solved using an exact three-step Taylor expansion time-stepping scheme. At time t=0, for all collocation points, the initial values of the solution v and its spatial derivatives are given as specific constants: - Function value: v = A = 4 - First partial derivatives: v_x = B and v_y = B (equal derivatives in x and y) - Second partial derivatives: v_xx = C = 5 and v_yy = C (equal second derivatives in x and y) After one full three-step time evolution with a time step of Δt = 1, the solution at any collocation point (xᵢ, yⱼ) is observed to be exactly 0, i.e., v(xᵢ, yⱼ, Δt) = 0. Question: To satisfy all of the above conditions, what must be the value of B², the square of the initial first derivative constant B? Round your answer to the nearest integer.	[ "Find the article title 'An algorithm for numerical solution of some nonlinear multi-dimensional parabolic partial differential equations'.", "Step 1: Establish the discrete framework. Define the operator F(v) = v_xx + v_yy - vv_x - vv_y. The three-step time evolution can be expressed as:\n1. v_intermediate1 = ...	3	mathematics	numerical_methods_for_differential_equations	micro-experiments
SGI_DeepResearch_0259	Consider a directed graph G with n=6 vertices and non-negative edge weights. The graph is subject to a fully dynamic sequence of vertex updates, and the algorithm maintains locally historical paths with smoothing applied. The initial graph has the following edge weight matrix (with +∞ for no edge): W₀ = [ [0, 3, +∞, 2, +∞, +∞], [+∞, 0, 4, +∞, 1, +∞], [+∞, +∞, 0, 3, 2, +∞], [+∞, +∞, +∞, 0, 5, 4], [+∞, +∞, +∞, +∞, 0, 3], [+∞, +∞, +∞, +∞, +∞, 0] ] Shortest paths are unique using extended weights with a tie-breaking function that assigns a unique identifier to each edge based on vertex indices, such as ID(u,v) = u + n*v. The initial set of historical paths includes all shortest paths. Update sequence Σ: - σ₁: update(vertex 2, new weights: w'{2,3}=2, w'{2,4}=1, w'{2,5}=3, w'{2,6}=2; other edges incident to 2 unchanged) - σ₂: update(vertex 1, new weights: w'{1,2}=4, w'{1,4}=3, w'{1,6}=5; other edges incident to 1 unchanged) - σ₃: update(vertex 3, new weights: w'{3,4}=2, w'{3,5}=4, w'{3,6}=3; other edges incident to 3 unchanged) - σ₄: update(vertex 4, new weights: w'{4,5}=2, w'{4,6}=1; other edges incident to 4 unchanged) - σ₅: update(vertex 5, new weights: w'_{5,6}=2; other edges incident to 5 unchanged) Smoothing is applied using the strategy described: after each update operation, a series of cleanup updates is triggered for vertices that were updated in the past, with timing based on exponential intervals relative to the current time. The current time range is up to t=5 (after σ₅). Cleanup updates overwrite edge weights without changing their actual values, only affecting the historical status of paths. After performing σ₅ and its associated smoothing burst, how many new locally historical paths are added for the pair of vertices (1,6)? Output the result as an integer. A path is considered new if it was not locally historical before σ₅ but becomes so after the update and smoothing. Locally historical paths are defined as paths where every proper subpath is a historical path, and historical paths are those that have been shortest paths at least once since the last vertex update on them. Additional details: - The graph has no negative-weight cycles. - The combinatorial properties of historical paths, including bounds on their numbers, should be considered.	[ "The paper 'A New Approach to Dynamic All Pairs Shortest Paths' introduces locally historical paths and smoothing. A path is locally historical if every proper subpath is historical, and a path is historical if it has been a shortest path at least once since the last vertex update on it.", "Initialize the graph w...	9	mathematics	shortest_path_planning	micro-experiments
SGI_DeepResearch_0265	In the THINGS-EEG dataset, the test set uses signal averaging to improve the signal-to-noise ratio (SNR). 80 repeated trials under the same image conditions are averaged. Given the SNR₀ for a single trial, assuming zero-mean Gaussian white noise and independent and identically distributed noise across trials, calculate how much the SNR improves after averaging over 80 trials compared to a single trial. Please give your answer to two decimal places.	[ "From paper \"Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion\", we found the test set averaging strategy: \"200 testing image conditions repeated 80 times\" and \"To improve signal-to-noise ratio, we averaged across the 4 EEG trials from the same image in the test set\".", "In the act...	19.03	neuroscience	visual_decoding	properties
SGI_DeepResearch_0269	A model is evaluated on three widely used EEG datasets: BCI Competition IV Dataset 2a, BCI Competition IV Dataset 2b, and SEED . Given these dataset types, which category of EEG data would such a model not be applicable to decode? Please answer with one acronym.	[ "Find the paper EEG Conformer: Convolutional Transformer for EEG Decoding and Visualization", "Extract information on the type of EEG paradigms each dataset represents.", "Identify that BCI IV 2a and 2b are used for motor imagery (MI) decoding tasks.", "Identify that SEED is used for affective or emotion reco...	ERP	neuroscience	motion_decoding	data
SGI_DeepResearch_0273	Large-scale pre-training has shown great potential to enhance models on downstream tasks in vision and language. Developing similar techniques for scalp electroencephalogram (EEG) is suitable since unlabelled data is plentiful. Meanwhile, various sampling channel selections and inherent structural and spatial information bring challenges and avenues to improve existing pre-training strategies further.Mapping various channel selections to a unified topology can break down boundaries between different EEG resources, facilitating cross-dataset EEG pretraining.What can be used to aggregate the temporal information extracted from the raw EEG?Please respond with a single phrase.	[ "Find paper Learning Topology-Agnostic EEG Representations with Geometry-Aware Modeling.", "Check the sentence:Large-scale pre-training has shown great potential to enhance models on downstream tasks in vision and language. Developing similar techniques for scalp electroencephalogram (EEG) is suitable since unlab...	DE feature	neuroscience	emotion_recognition	micro-experiments
SGI_DeepResearch_0277	In order to achieve the goal of reconstructing and interpreting neural circuit connectomics at synaptic resolution, existing electron microscopy techniques such as serial block-face scanning electron microscopy and focused ion beam scanning electron microscopy are constrained by efficiency and cannot meet the demand. In the process of neuron segmentation in electron microscopy datasets, auxiliary learning through predicting local shape descriptors can improve efficiency, which is very important for handling increasingly large datasets. Predicting local shape descriptors and other auxiliary learning tasks can enhance the segmentation accuracy and efficiency of affinity-based methods. When studying the accuracy and affinity of different methods, including local shape descriptors, on the ZEBRAFINCH, HEMI-BRAIN, and FIB-25 datasets, which metric is the key to ranking the various methods? Please do not use abbreviations in your answer.	[ "Find paper Local shape descriptors for neuron segmentation", "Find 'ZEBRAFINCH, HEMI-BRAIN, and FIB-25 datasets' and 'ranking'.", "Find out 'We find that the ranking of methods depends on the size of the evaluation ROI.'", "Search for the full name of 'ROI'", "Find out and conclude 'region of interest'" ]	region of interest	neuroscience	electron_microscopy_neuron_segmentation	data
SGI_DeepResearch_0284	Animals possess sophisticated control over their bodies, enabling diverse behavioral repertoires. However, how this control is achieved by the brain remains unclear. To advance understanding of this issue, models must be constructed that link control principles to the neural activity structures underlying behavior. Virtual Rodent, a biomechanical model driven by artificial neural networks within a physics simulator, employs deep reinforcement learning training while directly comparing results with neural activity recorded from real rats. This physics-based simulation, grounded in biomechanical realism, effectively explains neural activity patterns across behaviors and deeply integrates them with theoretical principles of motor control. How can a biological control system achieve the flexibility to control a complex object to perform different natural behaviors? Please answer using one phrase.	[ "Find paper A virtual rodent predicts the structure of neural activity across behaviors.", "Search for the keywords:Biological control systems,flexibility,natural behaviors.", "Find the sentence:Controlling a complex body to perform diverse natural behaviors requires a remarkable degree of flexibility in the un...	implementing internal models	neuroscience	neural_activity_and_behavior_prediction	properties
SGI_DeepResearch_0292	The design and discovery of high-temperature superconducting materials are crucial for advancing the development of next-generation electronic devices. The transition temperature (TC) of superconductors serves as a key performance parameter, typically predicted through electron-phonon coupling (EPC). The Bardeen-Cooper-Schrieffer (BCS) theory provides the theoretical framework for traditional superconductors. In a recent study, an efficient workflow for screening and predicting superconducting materials was developed by integrating BCS theory, density functional theory (DFT), and deep learning methods. As part of this process, the transition temperature TC of the material's VCs must be calculated based on given parameters: electron-phonon coupling parameter (λ): 1.2, logarithmic frequency (ωlog): 150 K, and μ parameter (coulomb pseudopotential): 0.09 (constant). According to BCS theory and the McMillan-Allen-Dynes formula:T_{c}=\frac{\omega_{\log }}{1.2} \exp \left[-\frac{1.04(1+\lambda)}{\lambda-\mu^{*}(1+0.62 \lambda)}\right], \quad \omega_{\log }=\exp \left(\frac{\int d \omega \frac{\alpha^{2} F(\omega)}{\omega} \ln \omega}{\int d \omega \frac{\alpha^{2} F(\omega)}{\omega}}\right) , the transition temperature TC (in Kelvin [K]) is retained with two decimal places. The final result can be directly provided.	[ "Find paper 'Designing high-TC superconductors with BCS-inspired screening, density functional theory, and deep-learning'.", "Position the introduction of the Bardeen-Cooper-Schrieffer (BCS) theory in the article.,", "The calculation results are used to predict the transition temperature (T_C) by McMillan-Allen...	13.95	physics	research_on_superconducting_mechanisms_discovery_of_superconducting_materials_and_process_optimization	properties
SGI_DeepResearch_0297	In the field of computational materials science, high-dimensional partial differential equations (PDEs) like the Allen-Cahn equation are pivotal for modeling phase transitions in 2D materials. In the context of high-dimensional PDEs, what is the exact memory reduction percentage achieved by the STDE method compared to traditional Taylor mode AD when solving the 100-dimensional Allen-Cahn equation, as demonstrated? Provide the value with 1 decimal place.	[ "From \"Stochastic Taylor Derivative Estimator: Efficient amortization for arbitrary differential operators\" Compares memory usage between STDE and traditional Taylor mode AD across dimensions.", "Identify the 100-dimensional case for Allen-Cahn equation in the document.", "Extract traditional Taylor mode AD m...	40.0	physics	computational_condensed_matter_physics_2	data
SGI_DeepResearch_0302	In the study of colloidal crystallization kinetics, the local bond order parameter \( q_6(i) \) quantifies the hexagonal symmetry of particle \( i \)'s neighborhood. For a colloidal system with particle diameter \( \sigma = \SI{2.2}{\micro\meter} \) and \( N_b(i) = 6 \) neighbors, if the spherical harmonics \( Y_{6,0}(\theta_{ij}, \phi_{ij}) \) for each neighbor \( j \) average to \( 0.3 \) and \( Y_{6,6}(\theta_{ij}, \phi_{ij}) \) average to \( 0.1 \), what is the value of \( q_6(i) \)? Provide the value with 3 decimal places.	[ "Find paper 'Visualizing kinetic pathways of homogeneous nucleation in colloidal crystallization'.", "The local bond order parameter \\( q_l(i) \\) is defined as \\( q_l(i) = \\left( \\frac{4\\pi}{2l+1} \\sum_{m=-l}^{l} \|q_{l,m}(i)\|^2 \\right)^{1/2} \\).", "For \\( l = 6 \\), compute \\( 2l + 1 = 13 \\), so \\(...	0.326	physics	research_on_soft_condensed_matter_physics_and_glass_transition_dynamics	properties
SGI_DeepResearch_0308	The experimental methodology for studying chaotic hysteresis in Chua's circuit is employs a precision Chua's circuit setup with calibrated instrumentation to investigate chaotic hysteresis through step-by-step DC voltage variation and frequency-dependent triangular wave analysis, quantifying hysteresis loops and identifying critical frequency thresholds via phase space trajectory monitoring and time series bifurcation analysis. In the Chua circuit experiment, what are the calculated time constants (in μs) for the RC networks formed by a 10.2 nF capacitor C1 and the equivalent resistance, the peak-to-peak voltage (in V) range of the hysteresis loop at 0.01 Hz driving frequency, and the critical frequency (in Hz) where chaotic behavior ceases? Output the results in two decimal places, one decimal place, and integer format respectively, separated by commas.	[ "From \"Experimental observation of chaotic hysteresis in Chua's circuit driven by slow voltage forcing\"", "Identify RC network components from experimental setup: C1=10.2 nF, R1=219Ω. Calculate time constant: τ=R1×C1=219×10.2×10−9=2.2338μs≈2.23μs.", "Voltage range determination: At 0.01 Hz triangular forcing,...	2.23, 3.2, 10	physics	chaotic_behavior_in_circuit_systems	properties
SGI_DeepResearch_0317	In experimental physics, particularly in setups involving sensitive magnetic field measurements, it is crucial to construct the apparatus from non-magnetic materials to avoid distorting the field. This includes structural components, sample holders, and optical elements. To ensure high magnetic field homogeneity for the optical widefield NMR microscope, several components in the vicinity of the magnet were specifically selected to be non-magnetic. Identify three distinct types of components that were chosen for this property. Provide the answer as three words, separated by commas.	[ "Find paper 'Optical widefield nuclear magnetic resonance microscopy'.", "Locate the 'Methods' section and the subsection 'Magnetic field homogeneity'.", "Read the paragraph describing the measures taken to ensure a homogeneous magnetic field.", "Identify the list of components selected to be non-magnetic.", ...	optics,cables,screws	physics	nuclear_magnetic_resonance_and_its_imaging_experiment	micro-experiments

Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows

Welcome to the official repository for the SGI-Bench! 👏

SGI Overview

Scientist-aligned benchmark for evaluating Scientific General Intelligence (SGI) across the full inquiry cycle: Deliberation, Conception, Action, and Perception. The benchmark spans 10 disciplines and more than 1,000 expert‑curated samples inspired by Science’s 125 Big Questions, with an agentic evaluation framework and multi‑metric protocol.

The Lite subset is designed for fast evaluation.

🆕 Latest News

🚩 Update (2025-12-22) We release SGI-Bench paper on arXiv.

🚩 Update (2025-12-19) SGI-Bench is adapted to VLMEvalKit and SciEvalKit, both of which are highly efficient and comprehensive evaluation toolkits.

🎤 Talk (2025-12-18) We are invited to give a talk on large language model evaluation at the AI Insight Talk jointly organized by OpenMMLab, Zhihu, and ModelScope.

🚩 Update (2025-12-12) We evaluate the newly released GPT-5.2-Pro on SGI-Bench.

👉 More News (Click to expand)

🚩 Update (2025-12-10) We update the paper PDF on the page.

🚩 Update (2025-12-03) We officially release the data and code of SGI-Bench.

🔬 What is Scientific General Intelligence (SGI)?

SGI denotes an AI system that can autonomously navigate the full, iterative cycle of scientific inquiry—Deliberation, Conception, Action, and Perception—with the versatility and proficiency of a human scientist. SGI‑Bench operationalizes this definition via four scientist‑aligned task families: scientific deep research, idea generation, dry/wet experiments, and multimodal experimental reasoning.

🎯 Framework & Tasks

SGI-Bench Pipeline

Deliberation (Scientific Deep Research): Multi‑hop retrieval, synthesis, and meta‑analysis style reasoning.
Conception (Idea Generation): Structured ideation and multi‑dimensional comparative evaluation.
Action (Dry/Wet Experiment): Code generation, lab protocol development and verification.
Perception (Experimental Reasoning): Process/observation/simulation/experiment/visualization image reasoning.

Grounded in the Practical Inquiry Model (PIM), SGI‑Bench treats science as an iterative cycle linking deliberation, conception, action and perception. Under this lens, SGI captures the capacity to integrate knowledge retrieval, idea formation, action execution, and interpretation into a unified loop of inquiry.

📂 Scientist‑Aligned Data Construction

Scientist-Aligned Data Construction

Raw Corpus: Expert‑curated texts/images across 10 domains, inspired by Science’s 125 Big Questions.
Question Construction: 100+ Master's and PhD holders with continuous expert‑in‑the‑loop review.
Data Cleaning: Rules + model checks + expert QA to ensure executability and unique answers.
Difficulty Filtering: Removes samples solved by >50% strong LLMs to maintain high challenge.

Result: High‑fidelity, scientist‑aligned tasks that are authentic, challenging, and broadly representative.

💯 Agentic Evaluation Framework

Agentic Evaluation Framework

Four Stages: Question Selection → Metric Customization → Predict & Eval → Report Generation
Tool Pool: Web search, PDF parser, Python interpreter, file reader, metric functions
Task Metrics: EM/SLA; Implementation Similarity; PassAll@k/SER; MCA/RV
Customizable: Add scientist‑aligned metrics (e.g., rigor, feasibility) on demand

This agent‑based stack formalizes scoring into traceable stages, improves reproducibility, mitigates evaluator–model coupling bias, and yields actionable, scientist‑aligned insights.

🚀 Test‑Time Reinforcement Learning (TTRL)

TTRL Training Dynamics

Objective: Address no‑ground‑truth idea generation by optimizing novelty at test time with online retrieval as a moving baseline.
Reward Design:
R = R_format + R_novelty
Enforce XML format and strict structure (e.g., <think>, <answer>); reward embedding dissimilarity from retrieved works, gated by thresholds.
Setup: GRPO on Qwen3‑8B (ms‑swift), G=8, high temperature, bfloat16, online retrieval n=4.
Dynamics: Format reward saturates quickly; novelty steadily increases. Average novelty improved from 49.36 → 62.06 without labels.

TTRL converts open‑ended ideation into measurable test‑time optimization and extends to multi‑objective rewards (rigor, feasibility, safety, cost).

🏆 Leaderboard Highlights

Model	Deep Research	Idea Generation	Dry Experiment	Wet Experiment	Experimental Reasoning	SGI-Score
Gemini-3-Pro 🥇	18.48	39.68	36.64	32.45	41.92	33.83
Claude-Sonnet-4.5 🥈	13.84	43.20	35.79	30.15	37.80	32.16
Qwen3-Max 🥉	15.38	39.83	33.21	33.62	37.80	31.97
GPT-4.1	11.32	36.49	34.32	36.63	38.49	31.45
GPT-5.2-Pro	15.72	55.03	28.04	17.50	39.18	31.09
GPT-5	14.47	55.40	29.89	16.31	38.14	30.84
o3	12.89	46.07	31.73	30.04	32.65	30.68
Claude-Opus-4.1	12.93	40.29	34.69	25.38	38.83	30.42
o4-mini	11.95	40.78	35.79	28.86	33.33	30.14
GPT-5.1	11.64	47.12	31.00	22.77	34.02	29.31
Grok-4	13.31	37.12	33.71	29.01	30.24	28.68
Qwen3-VL-235B-A22B	11.97	39.28	28.41	30.30	31.62	28.32
Gemini-2.5-Pro	15.09	39.95	22.51	22.05	41.24	28.17
Intern-S1	15.74	38.09	28.79	29.02	28.87	28.10
GPT-4o	7.86	35.95	26.94	31.31	32.30	26.87
Gemini-2.5-Flash	10.69	39.13	21.03	18.55	34.36	24.75
Llama-4-Scout	7.86	29.72	20.37	21.66	25.77	21.08
Qwen3-8B	8.18	35.78	18.45	9.96	23.37	19.15
Intern-S1-mini	11.06	36.04	16.97	12.42	16.84	18.67

🔥 Quick Start

git clone https://github.com/InternScience/SGI-Bench.git
cd SGI-Bench/evaluation

export OPENAI_API_KEY="xxxxx"
export OPENAI_BASE_URL="xxxxx"

conda create -n sgi python=3.13.7
conda activate sgi
pip install -r requirements.txt

📚 Task 1 Deep Research

conda activate sgi
python task_1_deep_research/step_1_get_answer.py gpt-5.2-pro
python task_1_deep_research/step_2_score.py gpt-5.2-pro

💡 Task 2 Idea Generation

Install the environment dependencies for evaluating idea generation.

conda create -n idea python=3.10.18
conda activate idea
pip install -r task_2_idea_generation/idea_generation_requirements.txt

Start the evaluation.

conda activate idea
python task_2_idea_generation/step_1_get_answer.py gpt-5.2-pro
python task_2_idea_generation/step_2_score.py gpt-5.2-pro

🖥️ Task 3.1 Dry Experiment (Code Generation)

Install the environment dependencies for running the dry experiment code.

conda create -n dryexp python=3.10.18
conda activate dryexp
pip install -r task_3_dry_experiment/dry_experiment_requirements.txt

Create code folder and initialize data (only need to run once).

conda activate sgi
python task_3_dry_experiment/step_1_build.py

Note: If some scripts time out during execution, please enter the corresponding folder and manually run the script to complete the data initialization.

Start the evaluation.

conda activate sgi
python task_3_dry_experiment/step_2_get_answer.py gpt-5.2-pro
python task_3_dry_experiment/step_3_run_code.py gpt-5.2-pro
python task_3_dry_experiment/step_4_score.py gpt-5.2-pro

🧪 Task 3.2 Wet Experiment (Lab Protocol)

conda activate sgi
python task_3_wet_experiment/step_1_get_answer.py gpt-5.2-pro
python task_3_wet_experiment/step_2_score.py gpt-5.2-pro

📊 Task 4 Experimental Reasoning

conda activate sgi
python task_4_experimental_reasoning/step_1_get_answer.py gpt-5.2-pro
python task_4_experimental_reasoning/step_2_score.py gpt-5.2-pro

💎 SGI-Score

conda activate sgi
python sgi_score.py gpt-5.2-pro

📬 Contact Us

💬 GitHub Issues: Please open an issue for bug reports or feature requests
📧 Email: xu_wanghan@sjtu.edu.cn
🤝 Community:

WeChat

📜 Citation

If you would like to cite our work, please use the following BibTeX.

@article{xu2025probing,
  title={Probing Scientific General Intelligence of LLMs with Scientist-Aligned Workflows},
  author={Xu, Wanghan and Zhou, Yuhao and Zhou, Yifan and Cao, Qinglong and Li, Shuo and Bu, Jia and Liu, Bo and Chen, Yixin and He, Xuming and Zhao, Xiangyu and others},
  journal={arXiv preprint arXiv:2512.16969},
  year={2025}
}