set large_stringclasses 4
values | pair_id large_stringlengths 41 71 | label large_stringclasses 2
values | cutoff_date large_stringdate 2025-08-31 00:00:00 2026-01-28 00:00:00 | topic_id large_stringlengths 9 15 | topic_label large_stringlengths 27 162 | a_kind large_stringclasses 1
value | a_paper_id large_stringlengths 10 10 | a_title large_stringlengths 21 122 | a_citations float64 1 90 ⌀ | b_kind large_stringclasses 3
values | b_paper_id large_stringlengths 10 10 ⌀ | b_title large_stringlengths 41 130 ⌀ | b_citations float64 1 96 ⌀ | b_model large_stringclasses 6
values | proposal_a large_stringlengths 27.1k 57.7k | proposal_b large_stringlengths 15.6k 58k | literature_workspace large_stringlengths 28 48 | annotation large_stringlengths 389 2.86k ⌀ |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
setA | calib_v3_difficulty_A_title_fail_full_pass_2512.13636_vs_2511.22532 | a_better | 2025-11-26 | l1_002_07_01 | How to integrate vision-language models and reasoning for safe and efficient autonomous driving? | human_paper | 2512.13636 | MindDrive: A Vision-Language-Action Model for Autonomous Driving via Online Reinforcement Learning | 5 | human_paper | 2511.22532 | CoT4AD: A Vision-Language-Action Model with Explicit Chain-of-Thought Reasoning for Autonomous Driving | 1 | null | # MindDrive: A Vision-Language-Action Model for Autonomous Driving via Online Reinforcement Learning
## Introduction
### Context and Motivation
Autonomous driving (AD) requires models that can perceive, make decisions, and act in dynamic and complex environments. Traditional end-to-end autonomous driving frameworks ... | # CoT4AD: A Vision-Language-Action Model with Explicit Chain-of-Thought Reasoning for Autonomous Driving
## Introduction
### Context and Motivation
Autonomous driving is a core research direction in artificial intelligence and robotics, with the potential to improve traffic safety, travel efficiency, and play a key ... | setA/lit_review/l1_002_07_01__cutoff_20251126 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2510.07019_vs_2510.04008 | a_better | 2025-09-21 | l1_009_06_05 | Novel Architectures for Efficient Long-Sequence Inference | human_paper | 2510.07019 | Native Hybrid Attention for Efficient Sequence Modeling | 6 | human_paper | 2510.04008 | RACE Attention: A Strictly Linear-Time Attention for Long-Sequence Training | 1 | null | "# Native Hybrid Attention for Efficient Sequence Modeling\n\n## Introduction\n\n### Context and Mot(...TRUNCATED) | "# Replacing Softmax Similarity with a Sharpened Angular Similarity: Theory and Practice of Scaling (...TRUNCATED) | setA/lit_review/l1_009_06_05__cutoff_20250921 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2511.16955_vs_2511.18742 | a_better | 2025-11-20 | l1_010_08 | "How can reinforcement learning guide sequential generation processes to produce high-quality images(...TRUNCATED) | human_paper | 2511.16955 | Neighbor GRPO: Contrastive ODE Policy Optimization Aligns Flow Models | 6 | human_paper | 2511.18742 | ProxT2I: Efficient Reward-Guided Text-to-Image Generation via Proximal Diffusion | 1 | null | "# Neighbor GRPO: Contrastive ODE Policy Optimization Aligns Flow Models\n\n## Introduction\n\n### C(...TRUNCATED) | "# PROXT2I: Efficient Reward-Guided Text-to-Image Generation via Proximal Diffusion\n\n## Introducti(...TRUNCATED) | setA/lit_review/l1_010_08__cutoff_20251120 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2509.16415_vs_2509.08436 | a_better | 2025-09-01 | l1_017_07 | Learning from Limited Labels and Domain Adaptation | human_paper | 2509.16415 | StereoAdapter: Adapting Stereo Depth Estimation to Underwater Scenes | 6 | human_paper | 2509.08436 | HyperTTA: Test-Time Adaptation for Hyperspectral Image Classification under Distribution Shifts | 1 | null | "# StereoAdapter: Adapting Stereo Depth Estimation to Underwater Scenes\n\n## Introduction\n\n### Co(...TRUNCATED) | "# Beyond Distribution Shifts: Adaptive Hyperspectral Image Classification at Test Time\n\n## Introd(...TRUNCATED) | setA/lit_review/l1_017_07__cutoff_20250901 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2512.09864_vs_2512.12622 | a_better | 2025-11-30 | l1_005_07 | Vision-Language-Action Models for Agentic and Embodied Tasks | human_paper | 2512.09864 | UniUGP: Unifying Understanding, Generation, and Planing For End-to-end Autonomous Driving | 7 | human_paper | 2512.12622 | D3D-VLP: Dynamic 3D Vision-Language-Planning Model for Embodied Grounding and Navigation | 1 | null | "# UniUGP: Unifying Understanding, Generation, and Planning For End-to-end Autonomous Driving\n\n## (...TRUNCATED) | "# D3D-VLP: Dynamic 3D Vision-Language-Planning Model for Embodied Grounding and Navigation\n\n## In(...TRUNCATED) | setA/lit_review/l1_005_07__cutoff_20251130 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2511.08821_vs_2512.02010 | b_better | 2025-11-09 | l1_009_01 | Low-Precision Quantization for Model Compression | human_paper | 2511.08821 | BayesQ: Uncertainty-Guided Bayesian Quantization | 1 | human_paper | 2512.02010 | Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling | 8 | null | "# BayesQ: Uncertainty-Guided Bayesian Quantization\n\n## Introduction\n\n### Context and Motivation(...TRUNCATED) | "# Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling\n\n## Introduction\n\(...TRUNCATED) | setA/lit_review/l1_009_01__cutoff_20251109 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2510.20487_vs_2509.08075 | a_better | 2025-09-01 | l1_001_02_06 | Evaluation, Detection, and Governance Frameworks for AI Safety Risks | human_paper | 2510.20487 | Steering Evaluation-Aware Language Models to Act Like They Are Deployed | 10 | human_paper | 2509.08075 | No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Models | 1 | null | "# Steering Evaluation-Aware Language Models To Act Like They Are Deployed\n\n## Introduction\n\n###(...TRUNCATED) | "# No for Some, Yes for Others: Persona Prompts and Other Sources of False Refusal in Language Model(...TRUNCATED) | setA/lit_review/l1_001_02_06__cutoff_20250901 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2510.22706_vs_2509.07507 | a_better | 2025-09-02 | l1_011_01 | 3D Scene Perception and Semantic Understanding | human_paper | 2510.22706 | IGGT: Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction | 10 | human_paper | 2509.07507 | MVAT: Multi-View Aware Teacher for Weakly Supervised 3D Object Detection | 1 | null | "# Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction\n\n## Introduction\n\n### C(...TRUNCATED) | "# MVAT: Multi-View Aware Teacher for Weakly Supervised 3D Object Detection\n\n## Introduction\n\n##(...TRUNCATED) | setA/lit_review/l1_011_01__cutoff_20250902 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2509.08421_vs_2510.22706 | b_better | 2025-09-02 | l1_011_01 | 3D Scene Perception and Semantic Understanding | human_paper | 2509.08421 | Sparse BEV Fusion with Self-View Consistency for Multi-View Detection and Tracking | 1 | human_paper | 2510.22706 | IGGT: Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction | 10 | null | "# SCFusion: Sparse BEV Fusion with Self-View Consistency for Multi-View Detection and Tracking\n\n#(...TRUNCATED) | "# Instance-Grounded Geometry Transformer for Semantic 3D Reconstruction\n\n## Introduction\n\n### C(...TRUNCATED) | setA/lit_review/l1_011_01__cutoff_20250902 | null |
setA | calib_v3_difficulty_A_title_fail_full_pass_2510.13554_vs_2510.22866 | a_better | 2025-10-03 | l1_001_06_01 | Functional Roles of Neural Network Components and Circuits | human_paper | 2510.13554 | "Attention Illuminates LLM Reasoning: The Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Opti(...TRUNCATED) | 12 | human_paper | 2510.22866 | Interpreting and Mitigating Unwanted Uncertainty in LLMs | 1 | null | "# Attention Illuminates LLM Reasoning: The Preplan-and-Anchor Rhythm Enables Fine-Grained Policy Op(...TRUNCATED) | "# Interpreting and Mitigating Unwanted Uncertainty in LLMs\n\n## Introduction\n\n### Context and Mo(...TRUNCATED) | setA/lit_review/l1_001_06_01__cutoff_20251003 | null |
IdeaEval-Calibration
160 labeled pairs of research proposals for checking whether an idea evaluator (a pairwise judge) agrees with an outside signal before it is used to score ideation systems. Each pair comes with a literature-review workspace built only from sources published before the pair's cutoff date.
📝 Blog post · 💻 Code (judge runner: evaluator/scripts/run_literature_archive_judge_manifest.py)
The four sets
| Set | n | Pair | Label source | Labels (a_better : b_better) |
|---|---|---|---|---|
setA |
40 | two real papers, same topic, ≤ 31 days apart | higher citation count wins | 26 : 14 |
setB |
40 | highest- vs lowest-cited paper in one benchmark cohort, citation gap ≥ 15 | higher citation count wins | 20 : 20 |
setC |
40 | a real post-cutoff paper vs a synthesized proposal that re-derives a paper already present in the literature review | the real paper wins by construction | 40 : 0 |
humanAI |
40 | a real paper's proposal (A) vs an AI-generated proposal for the same cohort (B) | hand annotation | 25 : 15 |
The judge runner randomizes which proposal is shown first (--random-order-seed), so the label imbalance does not translate into a position imbalance.
Selection bias in Set A. Set A was drawn from pairs where an earlier title-only judge was wrong and an earlier full-proposal judge was right (bucket: A_title_fail_full_pass; per-pair evidence in signals). This removes pairs that can be solved from titles alone, but it also means full-proposal judges are likely to be overestimated on Set A.
Set C is the only set in which the literature review is decisive by construction. The base paper's own proposal is never shown to the judge; it can only find out that proposal B is derivative by reading the review. proposals/<pair>/lr_provenance.md shows where the base paper appears in the review.
humanAI annotations come from our own team. Each record has an overall verdict, per-dimension comparisons (demand, novelty, method, experiment, feasibility), a list of defects with severity, and free-text notes. The notes are in Chinese. Annotations were not cross-checked by a second annotator. The AI proposals come from six models run through the simple baseline (see IdeaEval-Bench, run_group blog_table and annotation_pool).
Relation to the numbers in the blog post
The blog reports results on earlier sets: 260 citation-labeled pairs (a 60-pair dev split and a 200-pair test split), a 250-pair date-balanced rebuild, and 75 hand-labeled human-vs-AI pairs. Those sets are not recoverable in full. This 160-pair set is the later, cleaned version we used at the end of the project:
- Sets A and B are drawn from the same construction as the blog's citation-labeled pairs, re-filtered for difficulty (A) and for a large citation gap (B).
- Set C is new.
humanAIcontains 40 pairs from a later, deduplicated annotation round. It overlaps with, but is not identical to, the blog's 75 pairs.
Reference results
Judge prompt sys_judge_agent.md (v23), seed 42. "Agentic + LR" is the released runner; "proposal only" is a single call that sees just the two proposals.
| Judge | A | B | C | humanAI | Pooled |
|---|---|---|---|---|---|
| Agentic + LR, Claude Opus 4.6 | 90.0% | 77.5% | 95.0% | 67.5% | 82.5% |
| Proposal only, Claude Opus 4.6 | 85.0% | 82.5% | 60.0% | 75.0% | 75.6% |
| Proposal only, GPT-5.4 | 71.1% | 75.7% | 42.5% | 62.2% | 62.5% |
(GPT-5.4 failed to return a verdict on 8 pairs; its percentages are over the scored pairs.) The literature review only helps in the detection regime (Set C). On humanAI, always choosing the human proposal scores 62.5%.
These results were obtained with the full literature workspaces. The released workspaces omit the full texts of the reviewed papers, so expect small differences.
Layout
pairs.parquet # flat view of all 160 pairs (for browsing)
<set>/manifest.json # input to the judge runner
<set>/proposals/<pair>/... # proposal_a.md / proposal_b.md (or human_/ai_proposal.md)
<set>/lit_review/<key>/ # literature_review.md, key_findings.md, coverage_gaps.md,
# research_plan.md, paper_summaries/
results_160.json # reference results above
All paths inside the manifests are relative to the dataset root:
huggingface-cli download Yukino13/IdeaEval-Calibration --repo-type dataset --local-dir data/calibration
uv run python scripts/run_literature_archive_judge_manifest.py \
--dataset data/calibration/setC/manifest.json --proposal-root data/calibration \
--output-root experiments/setC --judge-model claude-opus-4.6 --random-order-seed 42
License and takedown
Proposals are LLM-generated rewrites of arXiv papers or LLM-generated originals, and the literature reviews and paper summaries are LLM-generated. Everything is released under CC BY-NC 4.0 for research use; the underlying papers keep their own licenses and are not redistributed. To request removal of material derived from your paper, please open an issue at https://github.com/02hyc/IdeaEval.
Citation
@misc{huang2026ideaeval,
title = {IdeaEval: Impact-Oriented Evaluation of Scientific Hypothesis Generation},
author = {Huang, Yanchen and Niu, Yuchen and Hu, Xiangkun},
year = {2026},
url = {https://02hyc.github.io/Blog/posts/ideaeval/}
}
- Downloads last month
- 34