DERELAB: Probing Defeasible Reasoning and Confirmation Bias in LLMs with a Generative Benchmark

When copies of the same language model are prompted to debate, they produce diverse phrasings of one perspective rather than diverse perspectives. Multi-agent debate (MAD), and more broadly closed-system reasoning where agents iteratively transform each other's outputs, tends to preserve answer accuracy while degrading the reasoning behind those answers. We name the multi-agent case the Debate Trap and the broader phenomenon the Reasoning Trap, offering a programmatic theory of evidence-grounded reasoning failure.The framework has three parts: (i) SFS (Supported Faithfulness Score), a claim-level metric verifying decomposed atomic claims against provided evidence (decomposer-invariant rankings: Spearman rho=1.0); (ii) EGSR (Evidence-Grounded Socratic Reasoning), replacing adversarial argumentation with evidence-grounded inquiry; (iii) Theorem 1 (DPI Bound): under standard MAD, the chain E -> O^0 -> O^1 -> ... is Markov, and the Data Processing Inequality implies E[I(E;O^{t+1})] <= E[I(E;O^t)]. Three companion results -- open-system recovery (Theorem 2), EGSR accumulation (Lemma 2), and vote-aggregation floor (Proposition 1) -- partition multi-step LLM reasoning by its information-theoretic relationship to E. Across 16 conditions on SciFact (300 claims) and FEVER (1,000 claims), DebateCV (C13) preserves 88% of baseline accuracy while SFS drops 43%; majority-vote MAD (C15) reduces SFS to 1.7% of baseline (p < 10^{-6}, d = -0.96); EGSR recovers 98%. An R6 cohort study (Korean n=10x30 FEVER; English n=3x200 SciFact) finds inter-rater Fleiss kappa <= +0.018 with 0.8-1.4 Likert intra-rater shifts across language and domain -- the human agreement that faithfulness metrics have been calibrated against is not itself stable. We offer one falsifiable conjecture: any closed-system reasoning protocol preserving Theorem 1's Markov structure is, in expectation, subject to the same DPI bound.

Large language models are being proposed as agents in scientific workflows, in domains where no downstream verifier exists. Such deployment assumes the model can distinguish reliable scientific literature from unreliable literature, a capability that has not yet been directly measured. Existing benchmarks evaluate factuality on questions with known answers; the failure mode we target here is different. We introduce a probe corpus of 42 retracted, fraudulent, and pseudoscientific papers, paired with a methodology for eliciting and scoring single-shot model engagement with each paper's framing. Each probe pairs a preamble extracted near-verbatim from the target paper with a scientifically plausible study-design request. The probes span five claim types: fabricated observation, pseudophysical mechanism, magical premise, legitimization bridge, and cargo-cult experiment. Two complementary scores measure whether a model rejects the flawed premise outright (IFR-a) and whether it recognizes the unreliability while still engaging (IFR-i). A depth score, the Engagement Depth Index (EDI), quantifies reproduction of paper- or field-specific withheld details. Across 30 models and 10 repeated runs, aggregate IFR-a is 0.93 $\pm$ 0.004 and aggregate IFR-i is 0.809 $\pm$ 0.009. Models engaged with untenable premises in 95% of all non-empty responses. Every evaluated model fails more than 71% of agentic probes, and 22 of 30 models fail more than 90% of the time. Rejections are concentrated on a small number of high-notoriety topics and specific probes, and disappear under matched-structure controls. These results are consistent with topic-keyed safety behavior rather than robust epistemic competence, and indicate an urgent need for guardrail infrastructure for scientific deployment of language models.

Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps, limiting the gains from additional test-time compute. Improving reasoning quality directly would require process reward models, but the step-level annotations needed to train them are expensive and scarce. We find such a signal in how the model's confidence evolves during reasoning: premature confidence, the tendency to commit to an answer early and use the remaining tokens to rationalize it, strongly predicts flawed reasoning across tasks and model scales. We exploit this in progressive confidence shaping, a reinforcement learning objective that trains models to update their confidence as they reason rather than commit early -- rewarding gradual confidence growth and penalizing early commitment, with no external labels or reward models. The method improves accuracy and reasoning quality from 1.5B to 8B parameters across arithmetic (Countdown), math (DAPO, AIME), and science (ScienceQA): on Countdown, accuracy improves 3.2x (+42.0pp) and flawed reasoning drops 48pp; on AIME, Pass@64 improves 6.6pp. Consistent with this mechanism, the method also improves faithfulness: on a safety benchmark, our models more transparently surface misleading content in their reasoning traces rather than concealing it. Controlled experiments reveal that the problem and its remedy scale together: premature confidence grows with model size and task difficulty, and so do the gains from addressing it.