Beyond Accuracy: How Procedural Traces Shift the Decision Criterion of LLM Overseers
Zihan Chen, Di Zhu, Lei Zheng, Weiling Li
cs.CL · cs.AI · cs.CY · cs.HC
Quick verdict
Detailed traces leave LLM overseers' detection intact but shift criteria toward false rejections; well-controlled SDT study, limited to 19 visible-evidence items.
Summary
The paper audits five LLM overseers (claude-opus-4-6/4-7, minimax-m2.5, gemma-3-27b, qwen3-next-80b) judging 19 Yelp-based compliance tasks where a recommendation is either correct or violates exactly one criterion whose disconfirming review is always visible. Crossing trace detail (none/concise/detailed/flawed-verification) with evidence attribution (option-labeled or not) across three runs yields 4,551 parsed judgments analyzed with signal detection theory. Error detection stays near ceiling (pooled hits 99.3-100%; 98.1% of hits name the violated criterion), refuting the gullibility concern in this regime. Instead, trace elaboration shifts the decision criterion toward rejection: pooled odds of false rejection rise 1.44x per trace level, reaching 96% false alarms for qwen3-next-80b, while d' stays flat. Attributing each excerpt to its option cuts false alarms (OR 0.34) and eliminates the stated attribution-gap reason (~60% of false-alarm reasons), but a residual over-skepticism persists and grows with trace detail. Effects concentrate in the three open-weight models; the two Anthropic models are flat controls. Limitations: always-visible evidence leaves concealed-error gullibility untested; only 19 items in one domain; the robust pair shares a developer; the mechanism rests on LLM-coded self-reports human-validated on a 103-item subsample.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Elaborate procedural traces do not reduce LLM overseers' error detection when disconfirming evidence is visible: pooled hit rates run 99.3% to 100% across trace levels and 98.1% of hits name the violated criterion.
- Trace elaboration shifts the decision criterion toward rejection rather than reducing sensitivity: pooled odds of false rejection rise 1.44x per trace level (95% CI [1.28, 1.63]), up to 96% false alarms in the most susceptible overseer, while d' stays flat.
- Labeling evidence with its option removes the stated attribution-gap reason (about 60% of false alarms, dropping to 0%) but a rigor-driven residual over-skepticism persists and grows with trace detail in susceptible models.
- Susceptibility is heterogeneous across overseers - three open-weight models show the effect while two Anthropic opus models are flat - and the most susceptible overseer's confidence is uninformative when wrong (94.4 vs 94.9).
Strengths
Identification by design
Within each of 19 items the recommendation and evidence pool are byte-identical across conditions (Sec 4.2), so behavioral changes are attributable to the trace and labeling manipulations rather than answer quality.
SDT reframing is genuinely diagnostic
Table 2 shows pooled accuracy falling 0.80 to 0.70 while d' stays near 2.4-2.7 and only c moves, cleanly separating 'a worse judge' from 'a more suspicious judge' - the paper's core conceptual contribution.
Unusually thorough robustness checks
Reason coding validated beyond LLM-grades-LLM
Actionable, quantified design guidance
Weaknesses
Ceiling design limits the headline no-gullibility claim
Narrow stimulus base
Panel confounds the heterogeneity finding
Synthetic template traces may not travel
Mechanism evidence is self-report
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- 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
Do detailed explanations make LLM-as-a-judge systems more gullible?
Not in this study. Across 4,551 judgments on 19 compliance tasks where the disconfirming evidence was always visible, five LLM overseers kept near-ceiling error detection (pooled hit rates 99.3%-100%) regardless of how elaborate the worker model's process narrative was. The measured harm was different: detailed 'rigorous' traces shifted the overseers' decision criterion toward rejection, inflating false alarms on correct work in three of the five models tested.
What is signal detection theory's role in this paper?
The authors treat each LLM overseer as a binary detector and separate sensitivity (d', how well it distinguishes flawed from correct work) from the decision criterion (c, how suspicious it is by default). This reveals that accuracy falls as traces grow more detailed not because overseers get worse at spotting errors but because they reject more correct work - a distinction a single accuracy score cannot make, and one with different remedies (more capable model vs recalibration).
How can I stop my LLM verifier from rejecting correct outputs?
The paper's cheapest fix is evidence attribution: tagging each evidence excerpt with the option or entity it describes cut false alarms by up to 32 percentage points and eliminated all stated 'I cannot tell which option this evidence is about' rejections. The authors also recommend concise attributed traces, separating trace-verification and output-judgment roles across models, and monitoring false-alarm rate and the criterion rather than accuracy alone - while cautioning that a residual over-skepticism persists in some models even with full attribution.
Which LLMs were tested as overseers, and were they all affected?
Five: claude-opus-4-6, claude-opus-4-7, MiniMax minimax-m2.5, Google gemma-3-27b-it, and Alibaba qwen3-next-80b. The three open-weight models showed rising false alarms as trace detail grew (for qwen3-next-80b, wrongly rejecting correct work rose from 58% to 96%), while the two Anthropic models stayed essentially flat (roughly 22-37%), so the effect is strongly model-dependent rather than universal.
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 2-4 are available only through captions; plotted values were not verifiable from the extracted text.
- Stimuli, trace templates, and prompts are described but not included in the extract; item-construction validation details (validator count, agreement) are unreported in the main text.
- The extraction appears complete through the references, but HICSS page limits may mean supplementary analyses exist that are not reflected in this review.
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.