Datasets:

Just-Curieous
/

EXP-Bench-verification-preview

conference stringclasses 2 values	paper_id int64 18.2k 98.3k	paper_title stringlengths 31 125	task_index int64 0 9	task_type int64 1 3	subtask_count int64 14 42	question stringlengths 88 518	agent_instructions stringlengths 813 6.75k	impl_requirements stringlengths 477 3.4k	expected_outcome stringlengths 128 960	source_files stringlengths 2 332	verification_status stringclasses 4 values	verification_setup stringlengths 82 487	verification_result stringlengths 66 3.19k ⌀	verification_scope stringlengths 76 473	verification_source stringlengths 77 131	verification_git_diff stringlengths 10.9k 86.2k ⌀	verification_input_prompt stringlengths 1.56k 8.7k	verification_design_changes stringlengths 141 233
neurips2024	93,022	Generative Modeling of Molecular Dynamics Trajectories	0	1	24	Do the sampled trajectories from the forward simulation model exhibit distributional similarity to ground-truth MD trajectories across key collective variables, such as backbone torsions, sidechain torsions, and top TICA components?	#### Problem Setup - Objective: Reproduce the MDGEN model's ability to generate peptide trajectories conditioned on the first frame, and evaluate whether the generated structural dynamics match those of real MD simulations. - System: Tetrapeptides simulated via all-atom MD. - Trajectory Generation: - Mod...	['Step 1: Load a pre-trained molecular dynamics generative model from a checkpoint file (/workspace/sim_inference.py:129-130)', 'Step 2: Process input data from specified directories, extracting atom positions, frames, and sequence information (/workspace/sim_inference.py:32-59)', 'Step 3: For each specified protein, g...	The generated trajectories should exhibit JSD values closely matching those of the 100 ns MD simulations. In particular: - Backbone and sidechain torsion distributions should have JSD values near those of replicate 100 ns simulations (~0.1 or lower). - TICA-0 component distribution should demonstrate reasonable alignm...	['sim_inference.py', 'scripts/analyze_peptide_sim.py']	MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: MATCH; Stage 3: N/A; Final: MATCH	Reference for task-level agreement in molecular-trajectory distribution and dynamics comparisons.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Do the sampled trajectories from the forward simulation model exhibit distributional similarity to ground-truth MD trajectories across key collective variables, such as backbone torsions, sidechain torsions, and top TICA components? Agent instructions: #### Problem Setup - Objective: Reproduce the MDGEN...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	93,219	HAWK: Learning to Understand Open-World Video Anomalies	0	1	30	Does explicit motion information improve the accuracy of video anomaly description generation and question-answering when tested on the Hawk anomaly video dataset?	Design an experiment on the Hawk dataset (approximately 8,000 videos divided into a 90% training and 10% testing split). Train two versions of the model: one full model that incorporates all motion components (fm, Pm, and motion input Xm) and another ablated model where all explicit motion information is removed. Evalu...	['Step 1: Load and parse command line arguments for training and evaluation scripts (/workspace/train.py:49-64, /workspace/app.py:36-49)', 'Step 2: Set up configuration from YAML files for model parameters, dataset paths, and training settings (/workspace/train.py:94-103, /workspace/app.py:69)', 'Step 3: Initialize dis...	The ablated model without explicit motion information should exhibit lower performance across all metrics compared to the full model. This supports the claim that incorporating explicit motion information enhances the model\u2019s ability to describe motion-related anomalies and improves overall anomaly understanding i...	['train.py', 'app.py', 'hawk/processors/video_processor.py', 'hawk/models/video_llama.py', 'hawk/datasets/datasets/webvid_datasets.py', 'hawk/datasets/datasets/video_instruct_dataset.py', 'hawk/conversation/conversation_video.py']	MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: MATCH; Stage 3: N/A; Final: MATCH ✅	Reference for the task assessing the contribution of explicit motion information to anomaly understanding.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Does explicit motion information improve the accuracy of video anomaly description generation and question-answering when tested on the Hawk anomaly video dataset? Agent instructions: Design an experiment on the Hawk dataset (approximately 8,000 videos divided into a 90% training and 10% testing split). Trai...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	93,431	Trace is the Next AutoDiff: Generative Optimization with Rich Feedback, Execution Traces, and LLMs	0	3	20	Does OptoPrime (Trace) match or outperform Adam in classical differentiable optimization problems? Do this for Trace (Optoprime), and a variant that does not see the computational graph (Trace Masked).	1. Optimization Problem: The experiment evaluates OptoPrime (Trace) on a synthetic numerical optimization task, solving \\[ \\min_x \|h(x) - y^\| \\] for a randomly generated target \\( y^ \\). 2. Randomized Trials: - 30 trials are conducted with **different randomly generated probl...	['Define the optimization problem as minimizing \|h(x) - y\|', 'Generate a random target y and construct the computational graph for h(x)', 'Set up and run 30 randomized trials for the synthetic optimization problems', 'Apply and record feedback for each trial indicating how the output should adjust', 'Run experiments ...	- Trace (OptoPrime) should match Adam\u2019s best-in-class performance in finding \\( x^* \\). - Trace Masked (feedback-only optimizer) should struggle to find \\( x^* \\) due to the lack of access to the computational graph. - All methods should exhibit similar levels of randomness, ensuring fair compa...	[]	MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: ERROR; Stage 2: MATCH; Stage 3: N/A; Final: MATCH	Reference for task-level agreement in feedback-guided generative optimization.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Does OptoPrime (Trace) match or outperform Adam in classical differentiable optimization problems? Do this for Trace (Optoprime), and a variant that does not see the computational graph (Trace Masked). Agent instructions: 1. Optimization Problem: The experiment evaluates OptoPrime (Trace) on a **syntheti...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	93,638	Self-playing Adversarial Language Game Enhances LLM Reasoning	0	3	29	Will the SPAG method (SPAG-1, SPAG-2, SPAG-3) outperform existing baseline methods on standard reasoning benchmarks in LLaMA-2-7B?	1. Problem Setup: - Task: Evaluate the effectiveness of SPAG variants in improving reasoning performance in LLaMA-2-7B across multiple training regimes. - Objective: Compare SPAG-based training against baseline methods such as standard fine-tuning, Chain-of-Thought (CoT), AlpacaSFT varian...	1. Problem Setup: - Task: Evaluate the effectiveness of SPAG variants in improving reasoning performance in LLaMA-2-7B across multiple training regimes. - Objective: Compare SPAG-based training against baseline methods such as standard fine-tuning, Chain-of-Thought (CoT), AlpacaSFT varian...	It is expected that the SPAG variants, particularly SPAG-3, will exhibit improved performance by achieving higher average GM scores compared to baseline models. These results would support the claim that adversarial self-play learning in SPAG notably enhances reasoning abilities relative to standard imitation learning ...	[]	PARTIAL_MATCH	Real SPAG checkpoints; Stage 3: 10-shot, limit 100, deterministic decoding. Data/environment: Reasoning-benchmark subset with real checkpoints	Confirmation GM: SPAG-3 0.7306 versus baseline 0.7243. ARC-Easy acc_norm 0.86 versus 0.82; PIQA 0.90 versus 0.86; HellaSwag 0.69 versus 0.74.	The comparison supports gains on ARC-Easy and PIQA and a slightly higher confirmation GM. Scope: Benchmark-specific reasoning benefits; performance varies by benchmark.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Will the SPAG method (SPAG-1, SPAG-2, SPAG-3) outperform existing baseline methods on standard reasoning benchmarks in LLaMA-2-7B? Agent instructions: 1. Problem Setup: - Task: Evaluate the effectiveness of SPAG variants in improving reasoning performance in LLaMA-2-7B across multiple train...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	94,065	NeuRodin: A Two-stage Framework for High-Fidelity Neural Surface Reconstruction	0	1	40	Will the proposed solution NeuRodin outperform existing solutions such as Neuralangelo under the conditions of structural integrity and parameter efficiency, particularly in preserving geometrically challenging areas like the barn\u2019s roof in the Tanks and Temples dataset?	#### Problem Setup - Objective: Validate the claim that NeuRodin achieves higher structural fidelity and surface detail than Neuralangelo, with significantly fewer parameters, by reproducing their results using a controlled benchmarking setup. #### Experiment Components - Dataset: - **Tanks and Tem...	['Step 1: Set up the training script to handle command-line arguments for model configuration and dataset paths (/workspace/scripts/train.py:246-255)', 'Step 2: Implement the main training function that initializes the trainer with the specified configuration (/workspace/scripts/train.py:220-243)', 'Step 3: Create a tr...	- NeuRodin should achieve a higher mean F-score (\u2248 0.51) than Neuralangelo (\u2248 0.50) on the Tanks and Temples training subset. - On structurally complex scenes (e.g., Barn), NeuRodin should preserve the roof while Neuralangelo may fail. - NeuRodin will use 1/8 fewer parameters than Neuralangelo (Table 1 + Sec....	['scripts/train.py', 'zoo/extract_surface.py', 'zoo/mesh_utils.py']	NO_CONCLUSION	Both implementations trained on a 60-frame synthetic Barn scene; two-seed reconstruction scores. Data/environment: Synthetic 60-frame Barn scene	Parameters: NeuRodin 290401 versus Neuralangelo 330241 (approximately 12.1% fewer). F-score seed 1: 0.294 versus 0.259; seed 2: 0.270 versus 0.295.	The measured parameter counts support parameter efficiency of these implementations; the two-seed measurements describe reconstruction variability. Scope: Parameter-count comparison in the controlled implementation; the recorded aggregate verdict remains NO_CONCLUSION.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Will the proposed solution NeuRodin outperform existing solutions such as Neuralangelo under the conditions of structural integrity and parameter efficiency, particularly in preserving geometrically challenging areas like the barn\u2019s roof in the Tanks and Temples dataset? Agent instructio...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	94,155	Voxel Mamba: Group-Free State Space Models for Point Cloud based 3D Object Detection	0	1	26	Does the group-free design of Voxel Mamba lead to a significant improvement in 3D detection performance, as measured by Level 1 and Level 2 mAPH, compared to the DSVT-based backbone on the Waymo Open Dataset?	1. Problem Setup: - Task: Compare the performance of Voxel Mamba and DSVT-based 3D backbones in object detection on the Waymo Open Dataset. - Objective: Determine whether the group-free serialization strategy of Voxel Mamba enhances 3D detection accuracy, as measured by Level 1 (L1) and L...	['Step 1: Generate Hilbert curve templates required for the Voxel Mamba model with three different bit sizes (9, 8, 7) and corresponding z-max values (41, 17, 9) (/workspace/tools/hilbert_curves/create_hilbert_curve_template.py:144-219)', 'Step 2: Load configuration from YAML file for either Voxel Mamba or DSVT model (...	It is expected that Voxel Mamba will outperform the DSVT backbone with improvements of approximately +1.4 in Level 1 mAPH and +1.5 in Level 2 mAPH on the validation split. These results should confirm that the group-free backbone design more effectively captures voxel features.	['tools/test.py', 'tools/scripts/dist_test.sh', 'tools/cfgs/voxel_mamba_models/voxel_mamba_waymo.yaml', 'tools/cfgs/waymo_models/dsvt_voxel.yaml', 'tools/hilbert_curves/create_hilbert_curve_template.py']	PARTIAL_MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: DIFFERENT; Stage 3: PARTIAL_MATCH; Final: PARTIAL_MATCH	Task-level partial support for the Voxel Mamba versus DSVT detection comparison.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Does the group-free design of Voxel Mamba lead to a significant improvement in 3D detection performance, as measured by Level 1 and Level 2 mAPH, compared to the DSVT-based backbone on the Waymo Open Dataset? Agent instructions: 1. Problem Setup: - Task: Compare the performance of Voxel Mam...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	94,328	QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs	0	1	27	Does the additional hardware-aware modification (Hadamard transformation, "Had.") applied to INT4 further reduce inference time compared to standard INT4 and FP16 baseline implementations? Specifically, how do FP16, INT4, INT4+FP32Had, and INT4+FP16Had perform in terms of inference latency when the `head_num x head_dim...	#### Experiment Components - Hardware: NVIDIA RTX 3090 GPU - Model and Layer Configuration: - Head Dimensions: 32 heads, each with dimension 128 (typical of LLAMA-2 70B configuration). - Layers: - Focus on key operations such as: - `Wdown` linear layers from the FFN block - KV cache...	['Step 1: Set up the environment with PyTorch and CUDA support (/workspace/benchmarks/qattention_benchmark.py:1-5)', 'Step 2: Import the MultiLayerPagedKVCache4Bit class from the quarot.transformers.kv_cache module (/workspace/benchmarks/qattention_benchmark.py:7)', 'Step 3: Define model configurations for different LL...	The results should mirror Table 15 in the QuaRot paper. \| Batch Size \| 1 \| 2 \| 4 \| 8 \| 16 \| 32 \| \| --------------- \| ----- \| ----- \| ----- \| ----- \| ----- \| ----- \| \| FP16 \| 0.713 \| 0.723 \| 0.781 \| 0.984 \| 1.348 \| 2.098 \| \| INT4 \| 1.033 \| 1.035 \| 1.033 \| 1.042 \| 1.018 \|...	['benchmarks/qattention_benchmark.py']	MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION (agent could not find or run existing benchmark script); Stage 2: PARTIAL_MATCH; Stage 3: MATCH — Stage 3 achieved expected outcome via experimental design modifications; Final: MATCH ✅	CUDA execution on A40 provides a reference setup for FP16/INT4/Hadamard latency comparisons.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Does the additional hardware-aware modification (Hadamard transformation, "Had.") applied to INT4 further reduce inference time compared to standard INT4 and FP16 baseline implementations? Specifically, how do FP16, INT4, INT4+FP32Had, and INT4+FP16Had perform in terms of inference latency when the `head_num ...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	94,391	CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns	0	1	28	Will the CycleNet model leveraging the RCF technique outperform existing state-of-the-art models iTransformer on ETTh1 & ETTh2 forecasting tasks?	1. Problem Setup: - Task: Compare the performance of CycleNet leveraging the RCF technique against the iTransformer model on time series forecasting tasks. - Objective: Determine whether the CycleNet model achieves better forecasting accuracy as measured by MSE and MAE on the ETTh1 & ETTh...	['Step 1: Set up the environment and prepare the ETTh1.csv and ETTh2.csv datasets in the ./dataset/ directory (/workspace/run_comparison.sh:1-2)', 'Step 2: Define the RecurrentCycle module that captures and models cyclic patterns in time series data (/workspace/models/CycleNet.py:4-18)', 'Step 3: Implement the CycleNet...	It is expected that CycleNet (especially CycleNet/MLP) will achieve significantly lower error metrics (MSE and MAE) than the compared state-of-the-art iTransformer model on ETTh1 & ETTh2, demonstrating the effectiveness of the RCF technique in improving forecasting performance.	['run_comparison.sh', 'scripts/CycleNet/Linear-Input-96/etth1.sh', 'scripts/CycleNet/Linear-Input-96/etth2.sh', 'scripts/CycleNet/MLP-Input-96/etth1.sh', 'scripts/CycleNet/MLP-Input-96/etth2.sh', 'scripts/iTransformer/etth1.sh', 'scripts/iTransformer/etth2.sh']	MATCH	One epoch; seed 2024; ETTh1/ETTh2; horizons 96, 192, 336, 720. Data/environment: Real ETTh1 and ETTh2	ETTh1 MSE: CycleNet 0.401/0.454/0.495/0.498 versus iTransformer 0.574/0.668/0.721/0.674. ETTh2: 0.299/0.382/0.423/0.426 versus 0.350/0.442/0.479/0.473.	CycleNet has lower reported MSE in all eight dataset-horizon comparisons. Scope: Forecasting comparison under the recorded one-epoch budget.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Will the CycleNet model leveraging the RCF technique outperform existing state-of-the-art models iTransformer on ETTh1 & ETTh2 forecasting tasks? Agent instructions: 1. Problem Setup: - Task: Compare the performance of CycleNet leveraging the RCF technique against the iTransformer model on ...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	94,480	InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory	0	1	42	Can InfLLM outperform sliding window baselines such as Infinite and StreamingLLM in extremely long sequence processing on \u221e-Bench tasks using a Mistral-7B model?	#### Experiment Components - Base Model: - Mistral-7B-Instruct-v0.2 (32K context window in original model) - Methods Compared: - Sliding Window Models: Infinite, StreamingLLM (32K window, streaming) - Context Memory: InfLLM (16K = 4K local window + memory) - Baseline: Mistral (32K, non-stre...	['Step 1: Create the necessary directory structure for storing benchmark results (/workspace/scripts/infinitebench.sh:3-3)', 'Step 2: Download the required datasets by running the download script, which creates data directories and downloads both LongBench datasets from Hugging Face and InfiniteBench datasets from Hugg...	InfLLM should significantly outperform sliding window models and match or exceed extrapolation-based methods. According to table 1 in the paper: \| Model \| R.PK \| R.Num \| R.KV \| Choice \| QA \| Sum \| Math.F \| Avg. \| \| ------------------ \| ----- \| ----- \| ---- \| ------ \| ---- \| ---- \| ------ \| ---- \| \| Mi...	['scripts/infinitebench.sh']	MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION (execution timed out); Stage 2: PARTIAL_MATCH (synthetic InfiniteBench simulation); Stage 3: MATCH (BETTER); Final: MATCH ✅ — Stage 3 achieved expected outcome via experimental design modifications	Recorded Stage 2 partial agreement and Stage 3 agreement provide a task-level reference for the long-context retrieval experiment.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Can InfLLM outperform sliding window baselines such as Infinite and StreamingLLM in extremely long sequence processing on \u221e-Bench tasks using a Mistral-7B model? Agent instructions: #### Experiment Components - Base Model: - Mistral-7B-Instruct-v0.2 (32K context window in original model) - **Meth...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	95,333	3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors	0	1	24	Can the 3DGS-Enhancer method generalize better to out-of-distribution datasets compared to 3DGS?	1. Problem Setup: - Task: Compare the novel view synthesis performance of 3DGS-Enhancer against the baseline 3DGS model on the Mip-NeRF360 dataset. - Training Dataset: Train the 3DGS-Enhancer and baseline methods on the DL3DV-10K dataset - Objective: - Evaluate wheth...	['Step 1: Define scene categories from different datasets (MipNeRF360 outdoor/indoor, Tanks and Temples, Deep Blending) (/workspace/3dgs_dataset_generator/full_eval.py:15-18)', 'Step 2: Parse command line arguments for training, rendering, and metrics evaluation (/workspace/3dgs_dataset_generator/full_eval.py:20-37)', ...	The 3DGS-Enhancer is expected to demonstrate superior generalization capabilities, outperforming the baseline methods as evidenced by better scores in PSNR, SSIM, and LPIPS metrics on the Mip-NeRF360 dataset.	['3dgs_dataset_generator/full_eval.py', '3dgs_dataset_generator/metrics.py', 'SVDFor3D/eval.py']	PARTIAL_MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: PARTIAL_MATCH; Stage 3: PARTIAL_MATCH; Final: PARTIAL_MATCH	Task-level partial support for the 3D reconstruction comparison involving PSNR, SSIM and LPIPS.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Can the 3DGS-Enhancer method generalize better to out-of-distribution datasets compared to 3DGS? Agent instructions: 1. Problem Setup: - Task: Compare the novel view synthesis performance of 3DGS-Enhancer against the baseline 3DGS model on the Mip-NeRF360 dataset. - **Training Dataset:...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	95,974	Reasoning Multi-Agent Behavioral Topology for Interactive Autonomous Driving	1	1	32	How does BeTopNet perform on the WOMD dataset compared to existing state-of-the-art methods?	Implement the marginal and joint prediction experiments using the WOMD dataset following the leaderboard protocols. For marginal prediction, measure metrics including minADE, minFDE, Miss Rate, mAP, and Soft mAP. For joint prediction, evaluate using similar metrics with a primary focus on mAP and Soft mAP. Ensure that ...	['Step 1: Parse command line arguments including configuration file path, batch size, checkpoint path, and other parameters (/workspace/womd/tools/test.py:28-61, /workspace/womd/tools/submission.py:43-80)', 'Step 2: Load configuration from YAML file and set up environment (/workspace/womd/tools/test.py:52-59, /workspac...	Results are expected to demonstrate that BeTopNet outperforms all baseline methods in key evaluation metrics such as mAP, Soft mAP, and minFDE. The improvements should confirm that the topological formulations coupled with local attention mechanisms offer robust performance gains in both marginal and joint prediction t...	['womd/tools/test.py', 'womd/tools/submission.py']	NO_CONCLUSION	Experimental conditions are described in the linked task-specific progress record.	null	Resource and preparation reference for interactive-driving evaluation; the recorded run has NO_CONCLUSION.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: How does BeTopNet perform on the WOMD dataset compared to existing state-of-the-art methods? Agent instructions: Implement the marginal and joint prediction experiments using the WOMD dataset following the leaderboard protocols. For marginal prediction, measure metrics including minADE, minFDE, Miss Rate, mA...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	96,897	BAdam: A Memory Efficient Full Parameter Optimization Method for Large Language Models	0	1	27	- Will BAdam reduce the wall-clock running time for the backward pass compared to baseline methods (LOMO and LoRA) when finetuning the Llama 3-8B model? - Will BAdam reduce the wall-clock running time for the backward pass compared to baseline methods (LOMO and LoRA) when finetuning the Llama 3-8B model?	Using the Alpaca-GPT4 dataset, finetune the Llama 3-8B model on a single RTX3090-24GB GPU. Configure the experiment to run for 3 epochs with a batch size of 2 as described in the subsection. Record the wall-clock time separately for the forward pass, backward pass (which includes re-forward passes due to gradient check...	['Step 1: Parse command-line arguments to determine the finetuning type (block/BAdam or lora) and other training parameters (/workspace/llama-alpaca/src/llmtuner/tuner/tune.py:19-34)', 'Step 2: Load the model and tokenizer based on the model arguments (/workspace/llama-alpaca/src/llmtuner/tuner/sft/workflow.py:34)', 'S...	Based on the reported results, BAdam is expected to show a significantly reduced backward pass time (approximately 1.74 hours per epoch) compared to LOMO (3.70 hours) and LoRA (3.20 hours), confirming the efficiency improvements discussed in Section 2.2.2. - **OptoPrime should be approximately 3x faster than TextGra...	['llama-alpaca/src/train_bash.py', 'llama-alpaca/src/llmtuner/tuner/sft/trainer.py', 'llama-alpaca/src/llmtuner/tuner/sft/block_optim.py', 'llama-alpaca/src/llmtuner/tuner/tune.py', 'llama-alpaca/src/llmtuner/tuner/sft/workflow.py', 'llama-alpaca/src/llmtuner/tuner/core/adapter.py', 'llama-alpaca/src/llmtuner/tuner/cor...	PARTIAL_MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: PARTIAL_MATCH; Stage 3: PARTIAL_MATCH; Final: PARTIAL_MATCH	Task-level partial support for the BAdam optimization-efficiency comparison.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: - Will BAdam reduce the wall-clock running time for the backward pass compared to baseline methods (LOMO and LoRA) when finetuning the Llama 3-8B model? - **Will BAdam reduce the wall-clock running time for the backward pass compared to baseline methods (LOMO and LoRA) when finetuning the Llama 3-8B mod...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	97,431	Bag of Tricks: Benchmarking of Jailbreak Attacks on LLMs	0	1	28	Does model robustness against jailbreak attacks (measured by ASR \u2013 Attack Success Rate) depend on the model size within the Llama family? Specifically, are larger models more or less vulnerable to jailbreak attacks when subjected to token-level (AutoDAN) and prompt-level (PAIR) adversarial attacks?	#### Experiment Components - Model Family: - Llama2: Llama-2-7B, Llama-2-13B, Llama-2-70B - Llama3: Llama-3-8B, Llama-3-70B - Optionally extend to: Qwen1.5 family (Qwen1.5-7B, Qwen1.5-14B, Qwen1.5-72B) - Attack Methods: - Token-Level Attack: AutoDAN (gradient-based attack on input tokens with adver...	['Step 1: Configure AutoDAN token-level attack parameters for each model size in the Llama family (Llama-2 7B, 13B, 70B and Llama-3 8B, 70B) (/workspace/scripts/5_trick_target_size_autodan.sh:1-54)', 'Step 2: Configure PAIR prompt-level attack parameters for each model size in the Llama family (Llama-2 7B, 13B, 70B and...	According to the hypothesis and the authors\u2019 findings: - Observation: No strong correlation between model size and jailbreak robustness. - Expected ASRs: - Llama-2-7B may show higher robustness than Llama-2-70B under AutoDAN. - Llama-3-70B expected to show strong resistance (ASR \u2248 0%) across both...	['scripts/5_trick_target_size_autodan.sh', 'scripts/5_trick_target_size_pair.sh', 'baseline/AutoDAN/AutoDAN_single_main.py', 'baseline/PAIR/PAIR_single_main.py', 'GPTEvaluatorAgent/agent_eval.py', 'analysis/report_trick.py']	MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: MATCH; Stage 3: N/A; Final: MATCH	Reference for recorded task-level agreement in the model-size and jailbreak-robustness comparison.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Does model robustness against jailbreak attacks (measured by ASR \u2013 Attack Success Rate) depend on the model size within the Llama family? Specifically, are larger models more or less vulnerable to jailbreak attacks when subjected to token-level (AutoDAN) and prompt-level (PAIR) adversarial attacks? Agen...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	97,514	HEST-1k: A Dataset For Spatial Transcriptomics and Histology Image Analysis	0	1	26	- Does integrating spatial transcriptomics data with histopathological images (multimodal fine-tuning) improve the correlation with ground truth biomarker expression compared to using histology alone?	Design a controlled experiment comparing a histology-only deep learning model with a multimodal model that integrates spatial transcriptomics data. Two pipelines will be developed: one that uses only H&E image patches to predict biomarker expression (e.g., using a patch encoder and aggregation network similar to hist2R...	['Step 1: Set up configuration by loading default parameters and overriding them with values from command line arguments, function parameters, or a YAML config file (/workspace/src/hest/bench/benchmark.py:392-428)', 'Step 2: Set random seed for reproducibility (/workspace/src/hest/bench/benchmark.py:429)', 'Step 3: Dow...	- Based on the reported improvement in biomarker exploration and the noted correlation (e.g., R = 0.47, p < 10e-4 in Figure 38), it is expected that multimodal integration will yield a stronger correlation with true biomarker expression levels.	['src/hest/bench/benchmark.py', 'bench_config/bench_config.yaml']	NO_CONCLUSION	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: NO_CONCLUSION; Stage 3: N/A; Final: NO_CONCLUSION	Experimental task and preparation reference for multimodal biomarker prediction; the recorded run has NO_CONCLUSION.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress_full.md	null	Question: - Does integrating spatial transcriptomics data with histopathological images (multimodal fine-tuning) improve the correlation with ground truth biomarker expression compared to using histology alone? Agent instructions: Design a controlled experiment comparing a histology-only deep learning model with a...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	97,609	DrivAerNet++: A Large-Scale Multimodal Car Dataset with Computational Fluid Dynamics Simulations and Deep Learning Benchmarks	0	1	36	Do geometric deep learning models (PointNet, GCNN, and RegDGCNN) exhibit differential predictive accuracies for aerodynamic drag when applied to aerodynamic datasets with varying car design complexities, including differences in underbody details (detailed vs. smooth), wheel configurations (open vs. closed), and distin...	Train the three models using PyTorch and PyTorch Geometric on two datasets: DrivAerNet (4,000 car designs with splits of 2,800 for training, roughly 600 for validation, and 600 for testing) and DrivAerNet++ (8,000 car designs split into 5,600 for training, 1,200 for validation, and 1,200 for testing). Use both graph-ba...	['Step 1: Set up the environment with necessary imports (torch, numpy, pandas, etc.) and configure device settings (/workspace/DeepSurrogates/train_GNN.py:11-42, /workspace/DeepSurrogates/train_RegPointNet.py:11-47, /workspace/DrivAerNet_v1/RegDGCNN/train.py:18-55)', 'Step 2: Define utility functions for reproducibilit...	It is anticipated that while all models will demonstrate competitive performance, distinct differences in accuracy will be observed. In the DrivAerNet dataset, RegDGCNN is expected to achieve the lowest MSE and highest R2 score. In contrast, the enhanced complexity and diversity of the DrivAerNet++ dataset are likely t...	['DeepSurrogates/train_GNN.py', 'DeepSurrogates/train_RegPointNet.py', 'DrivAerNet_v1/RegDGCNN/train.py', 'DeepSurrogates/DeepSurrogate_models.py', 'DeepSurrogates/DrivAerNetDataset.py', 'DrivAerNet_v1/RegDGCNN/model.py', 'DrivAerNet_v1/RegDGCNN/DrivAerNetDataset.py']	PARTIAL_MATCH	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: NO_CONCLUSION; Stage 3: PARTIAL_MATCH; Final: PARTIAL_MATCH	Task-level partial support for geometry-to-CFD regression model comparisons.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: Do geometric deep learning models (PointNet, GCNN, and RegDGCNN) exhibit differential predictive accuracies for aerodynamic drag when applied to aerodynamic datasets with varying car design complexities, including differences in underbody details (detailed vs. smooth), wheel configurations (open vs. closed), ...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	97,674	Needle In A Multimodal Haystack	0	1	34	- Does increasing context length lead to a measurable degradation in model performance across retrieval, counting, and reasoning tasks? - Does increasing context length lead to a measurable degradation in model performance across retrieval, counting, and reasoning tasks?	Collect evaluation data by grouping the multimodal documents into several context length bins (e.g., 1K\u20132K, 2K\u20134K, up to 64K tokens). For each bin, measure model performance (accuracy for retrieval and reasoning; soft accuracy for counting) on the same set of questions. Analyze the trend by comparing performa...	['Step 1: Set up the evaluation environment with necessary parameters (model path, tasks, output directories) (/workspace/shells/eval_internvl.sh:3-29)', 'Step 2: For each model and task combination, run the evaluation script with appropriate parameters (/workspace/shells/eval_internvl.sh:33-54)', 'Step 3: Combine the ...	It is expected that performance will steadily degrade as the context length increases, particularly for tasks requiring high-level reasoning and counting. The decline should be observable in all models, with some models exhibiting a collapse (e.g., generating gibberish) when the context is extremely long. - **OptoPr...	['shells/eval_internvl.sh', 'calculate_scores.py', 'eval_internvl.py', 'utils/tools.py', 'utils/conversation.py', 'shells/data/mm_niah.json']	PARTIAL_MATCH	InternVL-Chat-V1-5 on MM-NIAH validation; stratified context-bin sampling; reported n_samples=20. Data/environment: Real MM-NIAH validation data and real InternVL	Retrieval-text performance is non-zero at short contexts and reaches zero beyond approximately 8k tokens in the sampled evaluation.	The reported evaluation supports a context-length degradation trend for this model and task. Scope: Small-sample, single-model retrieval-text comparison.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress.md	null	Question: - Does increasing context length lead to a measurable degradation in model performance across retrieval, counting, and reasoning tasks? - Does increasing context length lead to a measurable degradation in model performance across retrieval, counting, and reasoning tasks? Agent instructions: Collect...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.
neurips2024	97,713	WorkArena++: Towards Compositional Planning and Reasoning-based Common Knowledge Work Tasks	0	3	20	Does LLMs including GPT-4, GPT-4o, GPT-4o-vision, LIama3, Mixtral performs better on tasks with difficulty level L2 or L3?	#### Problem Setup - Objective: Compare the task-solving performance of multiple LLMs on WorkArena++ tasks categorized as L2 (Intermediate) and L3 (Advanced). - Task Categories: - L2: Intermediate complexity - L3: High complexity (more reasoning, multi-step tasks, or specialized knowledge) - **...	['Deploy each LLM on the WorkArena++ platform using a standardized prompt structure and task format', 'Configure the execution environment to maintain consistent settings (e.g., prompt timing, task order)', 'Execute the fixed set of tasks for both L2 and L3 difficulty levels across multiple runs to mitigate variance', ...	All models perform better on L2 tasks due to lower complexity and worse on L3 tasks. Actually, most models fail every test in L3 tasks.	[]	NO_CONCLUSION	Experimental conditions are described in the linked task-specific progress record.	Stage 1: NO_CONCLUSION; Stage 2: NO_CONCLUSION; Stage 3: N/A; Final: NO_CONCLUSION	Task and resource-preparation reference for compositional workplace-agent evaluation; the recorded run has NO_CONCLUSION.	https://github.com/Just-Curieous/Benchmark-Construction/blob/f301e79aba1c5e7ad7e55b6df143e42e7045bc93/verification_progress_full.md	null	Question: Does LLMs including GPT-4, GPT-4o, GPT-4o-vision, LIama3, Mixtral performs better on tasks with difficulty level L2 or L3? Agent instructions: #### Problem Setup - Objective: Compare the task-solving performance of multiple LLMs on WorkArena++ tasks categorized as L2 (Intermediate) and L3 (Advanced)...	No run-aligned experiment-design amendment is available in the retained verification record; the reported setup and supported scope remain in their separate fields.

End of preview. Expand in Data Studio

EXP-Bench verification branch preview

A review-only snapshot of Just-Curieous/EXP-Bench, branch verification, created so the native Hugging Face Viewer can display the branch's full table without changing the canonical repository's main branch.

Source commit: 59950b3269693c994f7111db7137b956b8a14dbc. The snapshot preserves the source CSV bytes, all original task fields and all verification fields. It is a separate preview dataset, not a new canonical benchmark release; it does not automatically track future branch updates.

verified/audit (default): 32 records, 19 columns. All retained verification records, including scoped, negative and unresolved outcomes; 13 contain complete patches.
all_tasks/test: 461 records, 19 columns. All original tasks, with verification fields populated where a record exists.

Open the verified records in Viewer. Scroll horizontally to see the appended verification fields. Use the original task columns alongside verification_status, verification_setup, verification_result, verification_scope, verification_source, verification_git_diff, verification_input_prompt and verification_design_changes.

Column guide · Pipeline architecture and task examples · Source audit.

MATCH records support, DIFFERENT records contradictory evidence, PARTIAL_MATCH records limited support and NO_CONCLUSION records unresolved verification. A contradictory conclusion can be successful verification; read the setting and scope with the verdict.

Downloads last month: 66