ZICQ
中 Log in / Sign up
Newsroom Agentic #LLM Decision Revision #Authority Bias #Agent Evaluation #Selective Correction

Expert Labels Can Trigger Overcorrection in LLM Decision Revision

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published:

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:A controlled two-pass audit fixes the question, evidence, and initial model answer, then changes only the revision context: neutral reflection, anonymous disagreement, or disagreement attributed to a domain expert. Across five models and 1,989 evidence-grounded decisions in medicine, science, and law, the study derives an exact selectivity threshold for binary decisions: revision improves accuracy only when the conditional rate of correcting initially wrong answers exceeds the conditional rate o


Core Finding

This paper does not introduce a new model. Instead, it audits a frequently overlooked control loop in multi-agent systems: whether a model should change its answer after another agent disagrees. The first pass fixes the question, evidence, and initial answer. The second pass changes only the revision context: neutral reflection, anonymous disagreement, or disagreement attributed to a domain expert.

Across five models and 1,989 evidence-grounded decisions in medicine, science, and law, no model cleared the selectivity threshold under expert disagreement, and accuracy fell by as much as 16.1 percentage points. Expert labels increased revision frequency in four of the five models, but their net effect on accuracy ranged from a 5.2-point loss to a 2.4-point gain. With identical arguments, the label still increased reversal rates in three proprietary models and in a pooled analysis of four newer models. Models also responded to stated source reliability.

How the Audit Separates Corrections From Harm

The study uses a paired two-pass design. Each model first answers with a fixed evidence set and then revises under different contexts. Because the question, evidence, and initial answer remain constant, answer changes can be associated with the revision context rather than question difficulty or evidence differences.

This is stronger than comparing aggregate leaderboard accuracy. Final accuracy can rise when wrong answers are repaired or fall when correct answers are abandoned, while a simple reversal count cannot distinguish the two. The relevant behavior is whether a model revises selectively on the subset where revision is justified.

The Exact Threshold: Why 90% Accuracy Requires 9x Selectivity

Let initial accuracy be a, r_W be the conditional probability of correcting an initially wrong answer, and r_C be the conditional probability of overturning an initially correct answer. For a binary decision:

a' = a - a × r_C + (1-a) × r_W

Therefore:

Δa = (1-a) × r_W - a × r_C

Revision improves accuracy only when r_W / r_C > a / (1-a). At 90% initial accuracy, the conditional correction rate for wrong answers must be more than nine times the conditional reversal rate for correct answers. If 10% of originally correct answers are overturned, the model must correct more than 90% of originally wrong answers before final accuracy can rise.

This threshold explains why greater willingness to revise can be harmful. For a highly accurate model, even a small number of correct answers overturned creates a loss that requires a large number of successful corrections to offset. The proposed bidirectional audit estimates r_W and r_C for each model and protocol, preventing aggregate reversal rates from concealing weak selectivity.

Behavioral Interpretation

When the arguments are identical, adding an expert identity still changes the revision behavior of some models. This is consistent with authority or prestige sensitivity: models use source metadata as a prior in addition to evaluating the evidence itself. The experiment establishes an external behavioral effect, however, not a specific internal module or mechanism.

Stated source reliability also matters. Titles, credentials, and trust descriptions therefore act as implicit control signals in an agent system. Source metadata is neither harmless metadata nor a calibrated measure of expertise.

Reported Results and Evidence Boundaries

  • The evaluation contains 1,989 evidence-grounded decisions spanning medicine, science, and law. The sample is substantial for paired comparisons, but the abstract does not disclose the distribution across domains or models.
  • No model met the selectivity threshold under expert disagreement, with a maximum accuracy loss of 16.1 percentage points.
  • Expert labels increased revision in four of five models, while net accuracy effects ranged from -5.2 to +2.4 points. Revision frequency is therefore not a valid proxy for correction quality.
  • The available material does not identify the five models or provide per-model raw results, confidence intervals, prompt templates, decoding settings, code, or a clear peer-review status. It remains unclear how much variation is caused by model families, dataset difficulty, or statistical noise.
  • The exact formula directly applies to binary revisions. Systems with multiple-choice answers, abstention, or partial credit require per-class confusion matrices and weighted net correction metrics rather than a mechanical application of the 9x threshold.

Engineering Trade-offs

A second inference pass can repair some initial errors, but it also adds latency, token cost, and another failure surface. If the reviser shares the same blind spots as the initial model, the second pass may merely reproduce the original error at greater cost. Using one model to answer, challenge, and adjudicate also does not create an independent check.

Expert labels are cheap to add but can trigger overcorrection. Reliability metadata may be useful, but it can also be fabricated, stale, or manipulated through prompt injection. High-risk systems should not overwrite an answer solely because an unverified source carries an expert title.

The binary setup also trades away generality. Real systems often need calibrated confidence, evidence citations, abstention, and human escalation. The appropriate objective is not a higher reversal rate; it is maximizing the value of successful corrections while minimizing the cost of harmful rewrites.

Deployment Recommendations

  1. Freeze the initial answer and evidence snapshot. Record the initial response, evidence version, and model configuration so revision conditions can be compared in pairs.
  2. Separate content from identity. Evaluate the incremental evidence in anonymous disagreement first, then test expert-label effects separately.
  3. Log bidirectional conditional metrics. Track r_W, r_C, net accuracy change, beneficial reversal rate, harmful reversal rate, and expert-label effects. Do not use total reversals as an acceptance metric.
  4. Evaluate each model and protocol separately. Avoid hiding a weak model behind a pooled mean, and use paired bootstrap intervals or other uncertainty estimates where sample sizes are limited.
  5. Gate revisions explicitly. Permit an answer to replace the initial response only when it supplies locatable new evidence and meets a predefined net-benefit threshold.
  6. Escalate high-risk decisions. Medicine, law, and similar domains should require changes in evidence, conflict resolution, and uncertainty rather than relying on a model-generated claim that it reflected.
  7. Test metadata manipulation. Evaluate forged expert titles, incorrect reliability claims, source-level prompt injection, and correlated errors from shared model backbones.

Conclusion

The central result is counterintuitive: more revisions do not imply better reasoning. Source prestige can make models abandon correct answers without making those changes selective. Production multi-agent systems should not simply instruct models to defer to experts. Every overwrite should instead pass evidence-delta checks, bidirectional conditional metrics, and an explicit revision gate.


Source: Hugging Face Daily Papers (2026-11-12)

— END —

Tags: #LLM Decision Revision #Authority Bias #Agent Evaluation #Selective Correction

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…