User Feedback Provides a Unique Signal that LLMs Can not Detect
Shachar Don-Yehiya, Leshem Choshen, Omri Abend
cs.CL
Quick verdict
Argues user feedback is a potent yet judge-masked improvement signal; sound design, but one judge family and κ=0.24 ground truth limit scope.
Summary
This paper argues that naturally occurring user feedback is a highly effective improvement signal for LLMs, and that prior reports of its ineffectiveness stem from systematic bias in LLM-as-a-judge evaluation. The authors corrupt o3 responses to ArenaHard-v2.0 hard prompts with four error types (causality inversion, crucial omission, entity swap, logic-operator reversal), each paired with three simulated feedback levels, creating synthetic data with definitive ground truth; they complement this with real user feedback extracted from the ShareLM collection. Improver models (Gemini-3-Flash and Qwen3-8B) revise responses with and without feedback, and a Gemini-3.1-Pro judge measures issue resolution and pairwise preference. Feedback-informed revisions resolve 9–32% more issues on synthetic data and roughly 19–27% more on naturalistic data. Yet pairwise judges often prefer the without-feedback variant: on the subset where only the feedback-informed revision actually fixes the issue, the judge picks the correct response in just 54.6% (large improver, synthetic) and 34% (naturalistic) of cases. A correlated-failure analysis shows a model judges worst exactly on queries it cannot self-improve (72.2% vs 87.7% accuracy). Key limitations: no statistical significance testing, internal numeric inconsistencies (e.g., §5.1's 77.9%/42.7% vs Table 4's 80.8%/49.2%), reliance on a single judge family, and weak human agreement on the naturalistic ground truth (Cohen's κ=0.24). Code and data are released.
Review complete
Assessment at a glance
5
risk flags identified
● high 1
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Feedback-informed revisions resolve targeted issues at substantially higher rates than without-feedback revisions: +9–32% on synthetic data (Table 4) and roughly +19–27% on naturalistic data (Figure 1; abstract states 16–27%).
- Pairwise LLM judges systematically prefer the without-feedback revision even when it leaves the corruption unfixed (Tables 1, 6, 7).
- On the subset where only the feedback-informed revision succeeds, judges pick the correct response only 54.6%/80.9% (synthetic) and 34.0%/54.0% (naturalistic) of the time.
- Judge failures correlate with improver failures: a model judges least accurately exactly on queries it cannot self-improve (72.2% vs 87.7%, §5.4; self-judge 67.6% vs 80.8%, §6.1).
- Even weak feedback signals help with diminishing returns: problem-pointing feedback +6–17%, binary 'problem exists' +3–10% (Table 2).
Strengths
Controlled design with verifiable ground truth
Corrupting o3 answers to ArenaHard-v2.0 hard prompts via four prompted error types yields definitive ground truth, and human verification confirms 92.2% of hard-prompt corruptions are valid (§4, Appendix E); the invalid 59.4%-correct creative-writing category is correctly excluded from the main text.
Sharp diagnostic directly refuting prior work
The WFShouldWin subset analysis (§5.3) isolates cases where only the feedback-informed revision fixes the issue (N=205/574 synthetic, 215/274 naturalistic), giving a clean test that exposes the pairwise judge's anti-feedback preference (54.6%/80.9% synthetic; 34.0%/54.0% naturalistic, Table 8).
Novel correlated-failure finding
Layered human validation and honest cost reporting
Graded ablation of feedback strength
Weaknesses
No statistical inference anywhere
Circular ground truth, weakest exactly where headline numbers live
Internal numeric inconsistencies
Single judge family with entangled models
Overclaimed title and discussion
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper rebuts Liu et al. (2025)'s conclusion that naturally occurring user feedback is too noisy to improve LLMs. Three central claims: (1) feedback-informed revisions resolve targeted issues at substantially higher rates than revisions without feedback (+9–32% synthetic, Table 4; abstract reports 16–27% naturalistic, while Figure 1 computes 19.5–27.2pp); (2) pairwise LLM judges systematically prefer the without-feedback variant despite its lower resolution rate (Tables 1, 6, 7); (3) this reflects correlated failure — models judge worst exactly on queries they cannot self-improve (§5.4: 72.2% vs 87.7%; §6.1: self-judge 67.6% vs default 80.8%). The inference-only setting is acknowledged, yet the Discussion extrapolates to training and speculates the signal is 'not encoded in the model's weights'.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the main finding of this paper?
User feedback lets LLMs fix mistakes far more often than revising without feedback (9–32% more fixes on synthetic tests, roughly 19–27% more on real chat data), but LLM-as-a-judge evaluations often prefer the unfixed version — which is why earlier studies concluded feedback was a useless training signal.
How did the authors know which response was actually correct?
They deliberately planted known errors (reversed causality, omitted crucial details, swapped entities, flipped logic) into strong-model answers, so the response that undoes the planted error is verifiable ground truth; findings on real chat data were validated against this synthetic benchmark and human annotations.
How strong is the judge bias?
On cases where only the feedback-based revision fixed the planted error, the judge picked the correct answer just 54.6% of the time (synthetic, large improver) and 34% (real data). When a model judged its own outputs, accuracy fell further (67.6% vs 80.8% for a stronger external judge).
Is code or data available for this paper?
Yes — the authors released code on GitHub (shachardon/feedback-blindspot) and data on Hugging Face (shachardon/feedback-blindspot), though the experiments rely on preview versions of Gemini APIs, which may limit exact replication over time.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are extracted as bare numeric runs without labels; figure values (e.g., Figure 1 percentages) were read from caption text and corroborating tables.
- Appendix A cross-references appear garbled in extraction (e.g., hard-prompt pairwise results attributed to 'Table 3' where the content matches Table 11), so a few appendix attributions are uncertain.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.