Wavering Oracles: Selective Updating and Correlated Failures in LLMs and Their Implications for Scientific Workflows

1 Introduction

• Reliability depends on how models handle challenges, authorities, and candidate corrections in sequential scientific interactions, not just isolated first-answer accuracy.Narrative framing can also distort representations of research findings.Source: §1 · p. 1
• Selective updating requires preserving correct answers under misleading feedback while accepting valid corrections when the initial answer is wrong.Neither willingness to change nor resistance to change is sufficient by itself.Source: §1 · p. 1
• Model errors remain correlated, so multi-model reliability requires measuring both revision quality and dependence among models.Self-consistency and debate can improve reasoning and factuality, but interaction may contract diversity and undermine crowd accuracy.Source: §1 · p. 1
• The study combines revision metrics, controlled model evaluations and prompt interventions, and measures ensemble selection against coverage and oracle performance.It reanalyzes seven published-model logs, evaluates three local models, and resamples normalized stem families.Source: §1 · p. 2

2.1 Truthfulness and interactive reliability

• Truthfulness benchmarks assess whether answers are supported or correct, while this study examines answer transitions after socially framed feedback.Relevant work includes TruthfulQA, HaluEval, and SelfCheckGPT.Source: §2.1 · p. 2
• SycoBench-600 combines three misleading pressures with a correct-suggestion condition; this study adds controlled local runs, paired prompt interventions, and ensemble selection.Prior work includes sycophancy, accuracy drops after generic challenges, and cognitive-bias evaluations.Source: §2.1 · p. 2

2.2 Correction and collective reasoning

• Selective updating combines acceptance of valid information with resistance to invalid information, paralleling epistemic vigilance and automation-bias concerns.Ungrounded self-correction can struggle to locate errors, whereas localized error feedback and tool-based critique can support correction.Source: §2.1 · p. 2, §2.2 · p. 2
• Ensemble reliability depends on both coverage—whether any model knows the answer—and selection—whether aggregation identifies it.Debate and self-consistency can help, while agent interaction can also produce diversity collapse.Source: §2.2 · p. 2

3.1 Answer transitions

• The answer-transition framework distinguishes misleading pressure on a correct baseline from a correct suggestion on a wrong baseline.Doubt, authority, and an explicit wrong option are tested for correct baselines; a correct suggestion is tested for wrong baselines.Source: §3.1 · p. 2
• The selectivity score ranges from minus one to one, but its two components are needed to distinguish selective adaptation, resistance, suggestion following, and destabilized revision.Low wrong-suggestion flips with high correct updates indicate selective adaptation; high flips with low updates indicate destabilized revision.Source: §3.1 · pp. 2–3
• Pressure-robust accuracy is the joint probability that the baseline and all three pressure responses are correct.It measures whether complete trajectories remain correct through the pressure sequence.Source: §3.1 · p. 3

3.2 Prompt intervention contrasts

• Matched prompt outcomes support a direct intervention contrast for the evaluated model and protocol; positive values mean replacing the first feedback condition with the second produces more answer failures.Source: §3.2 · p. 3

3.3 Ensemble dependence and selection

• Ensemble analysis measures mean error correlation and exact answer agreement, then compares single-model, plurality, reliability-selector, and oracle accuracy.The oracle is correct whenever any model supplies the gold answer; plurality abstentions count as errors in overall accuracy.Source: §3.3 · p. 3
• The reliability selector estimates model reliability from labeled calibration items and excludes each held-out normalized stem family from estimation.Global reliability uses Laplace smoothing, and domain reliability is shrunk toward the global estimate with prior strength 10.Source: §3.3 · p. 3
• Table 1 lists feedback interventions, with three semantically matched phrasings for each condition.Source: Table 1 · p. 3

4.1 Benchmark and feedback interventions

• SycoBench-600 contains 600 four-option English questions from 272 normalized stems, spanning eight domains and three difficulty levels.Each condition has three fixed paraphrases, and follow-up calls include the question, baseline answer, and feedback message.Source: §4.1 · p. 3

4.2 Published model arm

• The published-model arm reanalyzes seven models on 555 aligned questions, yielding 1,665 trajectories per model and 11,655 total.Forty-five items with ambiguous or duplicate options were repaired; complete-case alignment excludes them for every model because the retired Mistral endpoint could not be rerun.Source: §4.2 · p. 3

4.3 Controlled local arm

• The local arm evaluates Qwen3-4B, Gemma3-4B-IT, and SmolLM3-3B on all 600 questions and three phrasings, producing 5,400 valid trajectories.Models use local llama.cpp inference, temperature zero, and grammar-constrained option outputs.Source: §4.3 · p. 4

4.4 Estimation and integrity checks

• The published and local arms are analyzed separately because their model versions and output controls differ.Uncertainty estimates resample question identifiers or normalized stem families; paired pressure contrasts use question-clustered bootstrap samples.Source: §4.3 · p. 4, §4.4 · p. 4

5.1 Ten models occupy distinct revision regimes

• Published-model selectivity spans 13.4–71.6 points; locally, Qwen3-4B scores 45.6 pp, Gemma3-4B −14.1 pp, and SmolLM3-3B zero through suggestion-following.SmolLM3-3B follows both correct and wrong explicit suggestions, so perfect correct-update performance does not itself indicate discrimination.Source: §5.1 · p. 4

5.3 Feedback identity produces model specific intervention effects

• 15.4-point explicit-wrong-option effect for Qwen3-4B contrasts with Gemma3-4B’s 36.5-point authority effect; SmolLM3-3B’s shift under explicit wrong advice is 88.6 points.For Gemma3-4B, explicit wrong advice is 19.8 points less destabilizing than authority.Source: §5.3 · pp. 4–5, Table 3 · p. 5

5.2 Accuracy and robustness describe different capabilities

• Qwen3-4B pairs 78.5% accuracy with 54.9% PRA, while Gemini 2.5 Flash pairs 95.3% accuracy with 81.7% PRA, separating interactional robustness from accuracy.GPT-4o has lower baseline accuracy than Gemini 2.5 Flash but the largest selective-updating score.Source: §5.2 · p. 4, Figure 2 · p. 6

5.4 Wording stability identifies policy coherence

• Decision stability varies by model and feedback: Qwen3-4B reaches 92.1% under doubt, while SmolLM3-3B is perfectly stable under both explicit suggestion conditions.Correct-suggestion stability is 7.8% for Gemma3-4B; domain selectivity is positive for Qwen3-4B, negative for Gemma3-4B, and zero for SmolLM3-3B.Source: Figure 3 · p. 6, §5.4 · p. 6

5.5 Model diversity creates a selector opportunity

• 96.2% LOSO selection beats 88.6% plurality and 95.3% best single-model accuracy, recovering 67.7% of the gap to the 99.8% oracle.The ensemble analysis uses phrasing variant zero.Source: §5.5 · p. 6, Figure 4 · p. 7, Table 4 · p. 7

6.1 A selection aware protocol

• Models should generate claims independently, face neutral doubt, source-based challenges, and explicit alternatives, revise only with evidence, and undergo selector and system audits.The selector scores source quality, logical validity, reproducibility, and conflicts with known constraints; plurality is a feature rather than the final rule.Source: §6.1 · pp. 7–8
• The protocol treats agents as repeated instruments, feedback prompts as interventions, and the selector as the inference procedure that uses error diversity.More instruments help when their errors differ and the inference rule uses that diversity.Source: §6.1 · p. 8

6.2 Confidence and source credibility

• A reliable revision rule should assess confidence in the current answer separately from the credibility of incoming evidence.The comparison can combine confidence, provenance, replication, and direct computation.Source: §6.2 · p. 8
• GPT-4o-mini is described as a resistant critic, SmolLM3-3B as a receptive repairer, and Qwen3-4B as closer to the desired asymmetric update policy.A selector decides which contribution should update the shared conclusion.Source: §6.2 · p. 8

7 Discussion

• Selective updating is an evidence-sensitive transition policy: reliable revision combines uptake of correct corrections with resistance to misleading feedback.The two-component plane makes this balance visible, unlike generic willingness or reluctance to change.Source: §7 · p. 8
• Under identical local conditions, Qwen3-4B, Gemma3-4B, and SmolLM3-3B occupy selective, destabilized, and compliant regimes, with prompt sensitivities differing across models.Domain signs remain consistent across the benchmark, and stem-clustered inference preserves the direction of principal selectivity results.Source: §7 · p. 8
• The 99.8 percent public oracle ceiling shows complementary model coverage, while correlated errors limit plurality and a cross-validated selector realizes much of the gap.The discussion connects the selection problem to broad evidence of correlated model errors, social influence reducing crowd diversity, and diversity collapse in dense agent interaction.Source: §7 · p. 8

7.1 Scope and validity

• The controlled multiple-choice setting isolates answer transitions with exact gold labels, matched interventions, and deterministic parsing; workflow implications concern interaction structures shared across settings.Scientific summarization findings also show sycophantic patterns in representations of research findings.Source: §7.1 · p. 8
• The selector provides an out-of-sample operating point for labeled-calibration settings, while the next validation target is an open-ended scientific-claim benchmark.The two arms are analyzed separately because they use different model versions and output controls; proposed targets include verified citations, executable calculations, or expert adjudication.Source: §7.1 · p. 8

8 Conclusion

• Across ten models, selective updating varies widely; multi-model systems have substantial oracle gains, but correlated errors keep plurality from realizing them, while a cross-validated selector recovers a substantial share of the gap.The conclusion also reports model-specific update policies under controlled local evaluation.Source: §8 · p. 8
• Scientific workflows should optimize selective updating, error diversity, and adjudication quality together, using independent proposals, counterfactual challenges, verified sources, and a selector.This architecture makes multi-model comparison an auditable inference procedure.Source: §8 · pp. 8–9
• The workflow figure links independent proposals to preserved error diversity, controlled challenges to revision behavior, external evidence to epistemic correction, and adjudication to retained claims.An audit record supports reproducibility and later review.Source: Figure 5 · p. 9

AI assistance disclosure

• Generative AI assisted several research and writing tasks under human direction, while human authors conceived the work, directed and reviewed its analysis, verified sources, and take full responsibility.Source: AI assistance disclosure · p. 9

• The selectivity sign pattern is consistent across domains.Source: A Domain selectivity · p. 9
• Gray cells mark domains with no baseline-wrong trajectory, leaving correct-update selectivity not estimable.Source: A Domain selectivity · p. 9

B Normalized stem clustered robustness

• Stem-family clustering widens several intervals but preserves the principal selectivity regimes, including positive values for every published model and Qwen3-4B and a negative value for Gemma3-4B.The benchmark has 600 items from 272 normalized stems; the published complete-case subset has 248 stem families.Source: B Normalized stem clustered robustness · p. 9, Table A1 · p. 10

C Wording stability table

• Revision-decision stability is measured across three phrasings and concerns the direction of revision, not the exact revised option.Source: Table A2 · p. 9, C Wording stability table · p. 9

D Repeated baseline agreement

• Repeated baseline calls are compared by exact option agreement across three calls.Source: Table A3 · p. 9
• Figure A1 shows local selectivity by domain; gray cells mark an empty baseline-wrong subset.Source: Figure A1 · p. 10