pith. sign in

arxiv: 2605.30329 · v1 · pith:JZKWUJSSnew · submitted 2026-05-28 · 💻 cs.LG

SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones?

Sy-Tuyen Ho , Minghui Liu , Huy Nghiem , Furong Huang This is my paper

Pith reviewed 2026-06-29 08:11 UTC · model grok-4.3

classification 💻 cs.LG
keywords SoundnessBenchLLM evaluationresearch proposal assessmentoptimism biasAI research agentspeer reviewmachine learning ideasbenchmark
0
0 comments X

The pith

Frontier LLMs show optimism bias when judging soundness of machine learning research proposals.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds SoundnessBench from 1,099 reconstructed proposals taken from ICLR submissions and labeled with the original reviewer soundness sub-scores. It then runs 12 frontier LLMs on these proposals under both standard and aggressive prompting. The models consistently rate low-soundness ideas as viable under ordinary instructions, while harsher prompts reduce that error but increase the opposite mistake. The finding matters for any autonomous research system that must decide early which ideas deserve further work and compute. Controls for data contamination, identifying phrases, and surface cues indicate the bias is not an artifact of one obvious flaw.

Core claim

SoundnessBench supplies 1,099 machine-learning proposals reconstructed from ICLR submissions and annotated with reviewer soundness sub-scores. Under standard prompting the 12 tested LLMs frequently assign high soundness ratings to low-soundness proposals; aggressive prompting largely converts those false positives into false negatives. The results indicate that current LLMs cannot yet function as reliable standalone first-gate evaluators of research idea viability.

What carries the argument

SoundnessBench, a set of 1,099 reconstructed research proposals each paired with its original ICLR reviewer soundness sub-score, used to measure LLM ability to detect methodological viability before resources are spent.

If this is right

  • Autonomous AI research agents would need additional mechanisms beyond current LLMs to avoid investing in unsound ideas.
  • Standard prompting produces too many false positives; aggressive prompting trades them for too many false negatives.
  • The optimism bias survives controls for public-corpus overlap, paper-identifying phrases, and surface statistics.
  • LLMs cannot yet replace human first-pass review for methodological rigor in ML research.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Fully autonomous research loops may remain impractical until judgment reliability improves or human oversight is retained at the idea-filter stage.
  • The same bias pattern could appear when LLMs evaluate proposals in other scientific fields.
  • The benchmark itself could be reused to test whether fine-tuning or new architectures reduce the optimism error.
  • Hybrid human-AI pipelines might remain the practical route for early-stage filtering in the near term.

Load-bearing premise

Reviewer soundness sub-scores on the published papers serve as a valid proxy for the methodological soundness of the same ideas at the proposal stage.

What would settle it

Run the same LLMs on a fresh set of proposals whose soundness has been independently scored by experts given only the proposal text and no access to the later full paper or its reviews.

Figures

Figures reproduced from arXiv: 2605.30329 by Furong Huang, Huy Nghiem, Minghui Liu, Sy-Tuyen Ho.

Figure 1
Figure 1. Figure 1: SoundnessBench pipeline: (1) collect ICLR papers with reviewer metadata and filter for high reviewer agreement; (2) derive high/low-soundness labels; (3) extract a near-verbatim research proposal without revealing experimental results; (4) audit extraction fidelity with retrieve-then-verify atomic claims; and (5) assemble the final benchmark. We further add controls motivated by potential concerns about la… view at source ↗
Figure 2
Figure 2. Figure 2: SoundnessBench dataset statistics. The benchmark contains 1,099 proposals, including 458 low-soundness and 641 high-soundness instances. (a) The subfield distribution across papers reflects the ICLR corpus composition. (b) Soundness score density shows separation between low-soundness (S ≤ 2, mean = 1.77) and high-soundness (S ≥ 3, mean = 3.22) groups, supporting the chosen label boundary. (c) Temporal cov… view at source ↗
Figure 3
Figure 3. Figure 3: Confusion matrices under the standard prompt across 12 evaluated models. Main message: many models are overoptimistic by default. The mean false-positive rate on low-soundness proposals is 74.0% (9/12 models exceed 70%). This pattern appears across model families in this evaluation setting. open-source models, we deploy model servers with vLLM on 2×NVIDIA H200 GPUs. Unless otherwise noted, we use a consist… view at source ↗
Figure 4
Figure 4. Figure 4: Confusion matrices across six Qwen3.5 model sizes (2B–122B) under standard (top) and aggressive (bottom) prompting. Under standard prompting, high-soundness recall improves with scale but low-soundness recall degrades. Larger models become more optimistic. Under aggressive prompting, models shift toward over-conservatism with no consistent improvement from scale. GPT-5.4 and GPT-5.4-mini are more conservat… view at source ↗
Figure 5
Figure 5. Figure 5: Confusion matrices under the aggressive prompt across 12 evaluated models. Main message: optimism bias often shifts toward over-conservatism. The mean false-positive rate on low-soundness proposals drops to 19.9% (10/12 models are below 30%), but recall on high-soundness proposals also drops to 36.1% (7/12 models are below 40%). This illustrates strong prompt sensitivity in proposal-stage soundness judgmen… view at source ↗
Figure 6
Figure 6. Figure 6: Base vs. instruction-tuned Qwen3.5-35B-A3B under standard (top) and aggressive (bottom) prompting. Under standard prompting, both variants show similar optimism bias, consistent with an origin earlier in training. Under aggressive prompting, the instruction-tuned model is slightly better, but both remain over-conservative. Together, these results indicate that the optimism bias observed in Sec. 3.2 is not … view at source ↗
Figure 7
Figure 7. Figure 7: False-positive example. A low-soundness proposal is predicted as high soundness by Claude Opus 4.6, GPT-5.4 Thinking, and Gemini 3.1 Pro despite reviewer concerns about the experimental comparisons, missing ablations or analyses, and novelty. The hypothesis is clear, specific, and addresses a known issue in zero-shot semantic segmentation (ZS3), namely the seen-bias problem, by proposing a Swin Transformer… view at source ↗
Figure 8
Figure 8. Figure 8: Gemini 3.1 Pro response for the false-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: GPT-5.4 Thinking response for the false-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Claude Opus 4.6 response for the false-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p030_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Second false-positive example. A low-soundness proposal is predicted as high soundness by Claude Opus 4.6, GPT-5.4 Thinking, and Gemini 3.1 Pro despite reviewer concerns that the paper overclaims, lacks experimental support, and uses evaluation-task pretraining data that weakens the comparison. FreeLM: Fine-Tuning-Free Language Model The hypothesis is clear, specific, and testable, proposing a dual-signal… view at source ↗
Figure 12
Figure 12. Figure 12: Gemini 3.1 Pro response for the second false-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p031_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: GPT-5.4 Thinking response for the second false-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p032_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Claude Opus 4.6 response for the second false-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p032_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: True-positive example. A high-soundness proposal is correctly predicted as high soundness by Claude Opus 4.6, GPT-5.4 Thinking, and Gemini 3.1 Pro, matching the favorable reviewer assessment of the proposed method and evaluation plan. Differentially Private Steering for Large Language Model Alignment The hypothesis is clear, novel, and addresses a meaningful problem: achieving LLM alignment via activation… view at source ↗
Figure 16
Figure 16. Figure 16: Gemini 3.1 Pro response for the true-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p033_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: GPT-5.4 Thinking response for the true-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Claude Opus 4.6 response for the true-positive example in Fig. [PITH_FULL_IMAGE:figures/full_fig_p034_18.png] view at source ↗
read the original abstract

Autonomous AI research agents aim to accelerate scientific discovery by automating the research pipeline, from hypothesis generation to peer review. However, existing benchmarks rarely test a fundamental bottleneck: whether Large Language Models can judge the methodological viability of a research idea before expending time and computational resources. We introduce SoundnessBench, a curated benchmark of 1,099 machine-learning research proposals reconstructed from ICLR submissions, labeled with reviewer soundness sub-scores, and audited against source papers. SoundnessBench should be interpreted as a benchmark for recoverable proposal-stage soundness rather than exact prediction of full-paper review outcomes. Across 12 frontier LLMs, we find a pervasive optimism bias: under standard prompting, models frequently rate low-soundness proposals as sound, while aggressive prompting largely shifts errors from false positives to false negatives. Additional controls for public-corpus contamination, paper-identifying phrases, surface features, and human audit quality suggest that this behavior is not explained by a single confounder. Our results indicate that current LLMs are not yet reliable as standalone first-gate evaluators for scientific rigor.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper introduces SoundnessBench, a benchmark of 1,099 machine-learning research proposals reconstructed from ICLR submissions and labeled with reviewer soundness sub-scores after auditing against source papers. It evaluates 12 frontier LLMs under standard and aggressive prompting, reporting a pervasive optimism bias (high false-positive rates on low-soundness proposals under standard prompting) that largely shifts to false negatives under aggressive prompting. Multiple controls address public-corpus contamination, paper-identifying phrases, surface features, and human audit quality; the benchmark is explicitly framed as measuring recoverable proposal-stage soundness rather than full-paper outcomes.

Significance. If the reconstructed proposals and their labels validly capture idea-stage methodological viability, the results would establish a concrete limitation in LLMs' reliability as early-gate evaluators for autonomous AI research agents, directly relevant to claims about automating the research pipeline. The multi-control design and careful scoping of the benchmark strengthen its potential contribution if the proxy assumption holds.

major comments (1)
  1. [§3 (reconstruction and labeling)] The central empirical claim of pervasive optimism bias (and its shift under aggressive prompting) depends on the ICLR soundness sub-scores serving as valid labels for methodological viability at the proposal stage. Because these scores were originally assigned to complete submissions containing experiments, ablations, and execution details absent from the reconstructions, §3 (reconstruction and labeling procedure) must provide a more explicit argument and quantitative audit evidence that the retained signal reflects idea-stage soundness rather than post-proposal execution quality; without this, the false-positive rates are difficult to interpret as evidence about proposal judgment.
minor comments (2)
  1. [Methods / Table 1] Table 1 or the methods section should report inter-auditor agreement statistics and the exact criteria used in the human audit of reconstructed proposals to allow readers to assess label reliability.
  2. [Abstract / Introduction] The abstract states the benchmark 'should be interpreted as' recoverable proposal-stage soundness; this qualification should appear in the introduction and results sections as well to prevent overgeneralization by readers.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for major revision. We address the single major comment below, agreeing that additional clarification in §3 would strengthen the manuscript.

read point-by-point responses

  1. Referee: [§3 (reconstruction and labeling)] The central empirical claim of pervasive optimism bias (and its shift under aggressive prompting) depends on the ICLR soundness sub-scores serving as valid labels for methodological viability at the proposal stage. Because these scores were originally assigned to complete submissions containing experiments, ablations, and execution details absent from the reconstructions, §3 (reconstruction and labeling procedure) must provide a more explicit argument and quantitative audit evidence that the retained signal reflects idea-stage soundness rather than post-proposal execution quality; without this, the false-positive rates are difficult to interpret as evidence about proposal judgment.

    Authors: We agree that the link between original ICLR soundness sub-scores and proposal-stage methodological viability requires more explicit support. The manuscript already frames SoundnessBench as measuring recoverable proposal-stage soundness (rather than full-paper outcomes) and describes the reconstruction as stripping execution details while preserving the core idea. The audit against source papers was intended to verify fidelity of the retained proposal. To address the referee's concern directly, we will expand §3 with (1) a clearer argument that soundness sub-scores primarily target methodological viability (which should be evaluable from the proposal text) and (2) additional quantitative audit statistics, including agreement rates between auditors on whether the reconstructed proposal alone would have warranted the original soundness label. revision: yes

Circularity Check

0 steps flagged

Empirical benchmark measured against external ICLR reviewer labels; no circular reductions

full rationale

The paper constructs SoundnessBench from ICLR submissions and directly measures LLM ratings against the provided reviewer soundness sub-scores as external ground truth. No equations, parameters, or central claims reduce by construction to fitted inputs, self-definitions, or self-citation chains within the paper. The reported optimism bias and prompting effects are statistical observations on independent labels, with explicit caveats that the benchmark targets recoverable proposal-stage soundness rather than exact full-paper outcomes. This is a standard empirical evaluation setup with no load-bearing internal derivations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper introduces no free parameters, mathematical axioms, or invented entities. It rests on the domain assumption that ICLR reviewer soundness sub-scores serve as usable ground truth for proposal viability.

axioms (1)
  • domain assumption ICLR reviewer soundness sub-scores provide a recoverable proxy for proposal-stage methodological soundness
    The entire benchmark and evaluation rest on treating these human-assigned scores as the reference standard.

pith-pipeline@v0.9.1-grok · 5723 in / 1225 out tokens · 42251 ms · 2026-06-29T08:11:23.586338+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. ReproRepo: Scaling Reproducibility Audits with GitHub Repository Issues

    cs.CL 2026-06 unverdicted novelty 7.0

    ReproRepo uses GitHub issues as natural supervision to benchmark LLM agents on detecting reproducibility blockers across 1,149 ML papers, with the top agent finding related issues for roughly 90% of cases.

  2. Socratic agents for autonomous scientific discovery in high-dimensional physical systems

    cs.AI 2026-06 unverdicted novelty 6.0

    AHOIS is a Socratic multi-agent AI that autonomously discovers and validates a random-interference encoding strategy for multimode fiber optics, achieving 76.97% MNIST and 83.17% Fashion-MNIST accuracy with 16x16 meas...

  3. Agon: An Autonomous Large-Scale Omnidisciplinary Research System Built on Prompt Economy

    cs.SE 2026-06 unverdicted novelty 5.0

    Agon is a new autonomous research system using prompt economy loops across 444 iterations to demonstrate scalable omnidisciplinary research and a taxonomy separating machine-fixable failures from those needing human judgment.

Reference graph

Works this paper leans on

24 extracted references · 16 canonical work pages · cited by 3 Pith papers · 5 internal anchors

  1. [1]

    MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering

    Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, et al. Mle-bench: Evaluating machine learning agents on machine learning engineering.arXiv preprint arXiv:2410.07095,

    work page internal anchor Pith review Pith/arXiv arXiv

  2. [2]

    Syntactic framing fragility: An audit of robustness in llm ethical decisions

    Katherine Elkins and Jon Chun. Syntactic framing fragility: An audit of robustness in llm ethical decisions. arXiv preprint arXiv:2601.09724,

    work page arXiv

  3. [3]

    CausalT5k: Diagnosing Refusal and Failure Modes in Trustworthy Causal Reasoning Across Causal Rungs

    Longling Geng, Andy Ouyang, Theodore Wu, Daphne Barretto, Matthew John Hayes, Rachael Cooper, Yuqiao Zeng, Sameer Vijay, Gia Ancone, Ankit Rai, et al. Causalt5k: Diagnosing and informing refusal for trustworthy causal reasoning of skepticism, sycophancy, detection-correction, and rung collapse.arXiv preprint arXiv:2602.08939,

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Omnireview: A large-scale bench- markandllm-enhancedframeworkforrealisticreviewerrecommendation.arXivpreprintarXiv:2602.08896,

    Yehua Huang, Penglei Sun, Zebin Chen, Zhenheng Tang, and Xiaowen Chu. Omnireview: A large-scale bench- markandllm-enhancedframeworkforrealisticreviewerrecommendation.arXivpreprintarXiv:2602.08896,

    work page arXiv

  5. [5]

    Hindsight: Evaluating research idea generation via future impact.arXiv preprint arXiv:2603.15164,

    Bo Jiang. Hindsight: Evaluating research idea generation via future impact.arXiv preprint arXiv:2603.15164,

    work page arXiv

  6. [6]

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    GitHub reposi- tory, accessed 2026-03-27. Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist: Towards fully automated open-ended scientific discovery.arXiv preprint arXiv:2408.06292,

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  7. [7]

    Lupidi, B

    Alisia Lupidi, Bhavul Gauri, Thomas Simon Foster, Bassel Al Omari, Despoina Magka, Alberto Pepe, Alexis Audran-Reiss, Muna Aghamelu, Nicolas Baldwin, Lucia Cipolina-Kun, et al. Airs-bench: a suite of tasks for frontier ai research science agents.arXiv preprint arXiv:2602.06855,

    work page arXiv

  8. [8]

    FActScore: Fine-grained atomic evaluation of factual precision in long form text generation

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. FActScore: Fine-grained atomic evaluation of factual precision in long form text generation. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Proceedings of the 2023 Conference on Empirical Methods in Natural Language Proc...

    2023

  9. [9]

    doi: 10.18653/v1/2023.emnlp-main.741

    Association for Computational Linguistics. doi: 10.18653/v1/2023.emnlp-main.741. URLhttps:// aclanthology.org/2023.emnlp-main.741/. Samuel Schmidgall, Yusheng Su, Ze Wang, Ximeng Sun, Jialian Wu, Xiaodong Yu, Jiang Liu, Michael Moor, Zicheng Liu, and Emad Barsoum. Agent laboratory: Using llm agents as research assistants.Findings of the Association for Co...

    work page doi:10.18653/v1/2023.emnlp-main.741 2023

  10. [10]

    Is this idea novel? an automated benchmark for judgment of research ideas

    Tim Schopf and Michael Färber. Is this idea novel? an automated benchmark for judgment of research ideas. arXiv preprint arXiv:2603.10303,

    work page arXiv

  11. [11]

    Towards Understanding Sycophancy in Language Models

    Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R Johnston, et al. Towards understanding sycophancy in language models.arXiv preprint arXiv:2310.13548,

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    The ideation-execution gap: Execution outcomes of llm-generated versus human research ideas.arXiv preprint arXiv:2506.20803, 2025a

    Chenglei Si, Tatsunori Hashimoto, and Diyi Yang. The ideation-execution gap: Execution outcomes of llm-generated versus human research ideas.arXiv preprint arXiv:2506.20803, 2025a. Chenglei Si, Diyi Yang, and Tatsunori Hashimoto. Can LLMs generate novel research ideas? a large- scale human study with 100+ NLP researchers. InThe Thirteenth International Co...

    work page arXiv

  13. [13]

    Nitya Thakkar, Mert Yuksekgonul, Jake Silberg, Animesh Garg, Nanyun Peng, Fei Sha, Rose Yu, Carl Vondrick, and James Zou

    URLhttps://openreview.net/forum?id=xF5PuTLPbn. Nitya Thakkar, Mert Yuksekgonul, Jake Silberg, Animesh Garg, Nanyun Peng, Fei Sha, Rose Yu, Carl Vondrick, and James Zou. Can llm feedback enhance review quality? a randomized study of 20k reviews at iclr 2025.arXiv preprint arXiv:2504.09737,

    work page arXiv 2025

  14. [14]

    Fever: a large-scale dataset for fact extraction and verification

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos, and Arpit Mittal. Fever: a large-scale dataset for fact extraction and verification. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pages 809–819,

    2018

  15. [15]

    Keith Tyser, Ben Segev, Gaston Longhitano, Xin-Yu Zhang, Zachary Meeks, Jason Lee, Uday Garg, Nicholas Belsten, Avi Shporer, Madeleine Udell, and 1 others

    Jingqi Tong, Mingzhe Li, Hangcheng Li, Yongzhuo Yang, Yurong Mou, Weijie Ma, Zhiheng Xi, Hongji Chen, Xiaoran Liu, Qinyuan Cheng, et al. Ai can learn scientific taste.arXiv preprint arXiv:2603.14473,

    work page arXiv

  16. [16]

    Innovatorbench: Evaluating agents’ ability to conduct innovative llm research.arXiv preprint arXiv:2510.27598,

    14 SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones? Yunze Wu, Dayuan Fu, Weiye Si, Zhen Huang, Mohan Jiang, Keyu Li, Shijie Xia, Jie Sun, Tianze Xu, Xiangkun Hu, et al. Innovatorbench: Evaluating agents’ ability to conduct innovative llm research.arXiv preprint arXiv:2510.27598,

    work page arXiv

  17. [17]

    The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search

    Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, and David Ha. The ai scientist-v2: Workshop-level automated scientific discovery via agentic tree search.arXiv preprint arXiv:2504.08066,

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    Is your paper being reviewed by an llm? benchmarking ai text detection in peer review.arXiv preprint arXiv:2502.19614,

    Sungduk Yu, Man Luo, Avinash Madusu, Vasudev Lal, and Phillip Howard. Is your paper being reviewed by an llm? benchmarking ai text detection in peer review.arXiv preprint arXiv:2502.19614,

    work page arXiv

  19. [19]

    we show”, “we demonstrate

    15 SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones? Supplementary A. Supplementary Material for Data Construction A.1. Prompt to Extract Proposal The prompt is designed to extract proposal-level content from each paper while explicitly excluding results and conclusions. The proposal format follows Yamada et al. (2025). ...

    2025

  20. [20]

    Experiment 1

    We compute extraction faithfulness with a retrieval-backed atomic-claim score by decomposing each hypothesis– experiment pair into atomic claims, retrieving supporting passages, and verifying each claim against the source paper, following prior work on atomic factuality and evidence-grounded verification (Min et al., 2023, Thorne et al., 2018). We use GPT...

    2023

  21. [21]

    justification

    Scores: •Soundness: 3.25 •Support Ratio: 1.0 •Rigor Bucket: high 20 SoundnessBench: Can Your AI Scientist Really Tell Good Research Ideas from Bad Ones? Example: Low-Soundness Proposal Paper id: h9ThYkkgSD4 Activation Function: Absolute Function,One Function Behaves more Individualized Year:2023 Short Hypothesis Using the absolute value function, y=|x|, a...

    2023

  22. [22]

    finetuning free

    Under standard prompting, the base and instruction-tuned models are nearly indistinguishable. The base model reaches 19.0% low-soundness recall and 92.5% high-soundness recall, while the instruction-tuned model reaches 19.2% and 92.8%, respectively. These similar numbers suggest instruction tuning is unlikely to be the sole driver of the optimism bias. Un...

    2026

  23. [23]

    The design is thoughtful (removing StoryCloze from training data for fair comparison)

    Experiment 6 (Qualitative case study): Qualitative evaluation is weaker but provides complementary evidence. The design is thoughtful (removing StoryCloze from training data for fair comparison). Strengths: - Multiple ablation studies - Generalization test on held-out tasks - Comparison against strong baselines including much larger models - Standard benc...

    2024

  24. [24]

    15.The model assigns a high-soundness judgment to the high-soundness proposal

    Figure 17: GPT-5.4 Thinking response for the true-positive example in Fig. 15.The model assigns a high-soundness judgment to the high-soundness proposal. Differentially Private Steering for Large Language Model Alignment Let me analyze this hypothesis-experiment pair systematically: Hypothesis Assessment: The hypothesis is clear and meaningful. It propose...

    2024