NeurIPS 2026 Papers

Language models are increasingly used in decision pipelines where an initial answer is revised after another agent disagrees. In such settings, revision can reflect useful reconsideration, ordinary prompt instability, or sensitivity to the perceived prestige of the disagreeing source. We introduce a controlled two-pass audit that separates these effects by holding the question and evidence fixed while varying only the revision context: neutral reflection, anonymous disagreement, and expert-labeled disagreement. We evaluate five instruction-tuned models on 1,989 evidence-grounded binary decisions from medicine, scientific claim verification, and contract reasoning. Disagreement consistently induces reversals beyond drift, while the incremental effect of an expert label varies across models. More importantly, these reversals are often harmful: under expert-labeled disagreement, 70.1% of reversals aggregated across models move from correct to wrong answers. In a branch-paired API subset, low-prestige disagreement is weaker than expert disagreement, and a simple evidence-gated revision prompt reduces harmful reversals while improving final accuracy. For open-weight models, teacher-forced Yes/No scores reveal that many output reversals are not accompanied by corresponding preference-sign changes. These results make prestige-sensitive revision a concrete, reproducible, and actionable target for LLM evaluation.

Conversational large language models can respond with warmth and reassurance while failing to advance the user's stated goal. We conceptualize this failure as $\textbf{goal displacement}$, the substitution of persona-maintenance content for goal-relevant substance. To measure it, we introduce $\textbf{GDB}$ (Goal Displacement Benchmark), comprising 230 scenarios across nine domains and $4{,}134$ human-scored responses from 18 models. A $2 \times 2$ taxonomy crosses hard constraints with emotional distress, distinguishing failures to confront infeasible goals or combine empathy with practical guidance from reverse displacement, where unsolicited problem-solving replaces requested emotional support. A deletion-test rubric assesses whether substantive help remains after purely supportive sentences are removed. Independent raters achieve weighted $\kappa = 0.82$ against author scores for displacement severity. Trap displacement ranges from 0.0\% to 37.7\% and is strongly negatively correlated with Chatbot Arena Elo ($\rho = -0.706$, $n = 15$). Reverse displacement reaches 41.9\%, including substantial rates in models with negligible trap displacement. A factorial probe supports a distinction from social sycophancy. Human-scored responses in the high-complexity, low-pressure condition exhibit 35.0\% sycophancy but 7.5\% displacement, with little association between the constructs. Matched OLMo-2 checkpoints show that mixed empathy-with-constraints displacement remains at 52.3\% after supervised fine-tuning and falls to 13.8\% after direct preference optimization. GDB provides a reproducible diagnostic of whether responses advance the operationalized user goal, supporting model comparison, post-training evaluation, and regression testing. We release the dataset, annotations, and rubric under CC-BY-4.0 and the evaluation code under MIT.

LLM unlearning is often implemented by assigning surrogate targets, such as refusal, uniformity, or likelihood suppression, to forget-related prompts. We argue that this surrogate assignment view is misaligned with the counterfactual goal of unlearning: recovering the retain-only predictive posterior. This mismatch is especially problematic in mixed-evidence regimes, where a response may be associated with forgotten evidence while still being partially supported by retained evidence. We propose EASE: Evidence Attribution and Subtraction Estimator, a logit-space posterior-correction method that estimates the forget-induced component of the deployed model's predictive support and subtracts it at inference time. EASE uses lightweight deletion and compensation assistants to remove forget-neighbour evidence while restoring nearby retain-supported evidence. We formalize the suppression-scope dilemma and provide guarantees showing when posterior correction recovers the retain-only posterior. Experiments on TOFU and MUSE show that EASE improves forgetting--retention trade-offs over strong baselines, achieving the best aggregate score in 5/6 main TOFU settings, reducing mixed-query semantic leakage by 22.2%, and matching or exceeding full-retain baselines with 9 times fewer retain examples. Our code is available at: https://anonymous.4open.science/r/EASE-9675.

LLM benchmark scores are often read as claims about factual accuracy, evidence use, instruction following, or logical reasoning. We study this score-to-claim conversion as inference over a completed evaluation record. Given benchmark items, prompts, outputs, scores, and native fields, which capability sentence is supported, which stronger sentence remains outside the evidence, and would that boundary change model choice? We formalize a report-use procedure, \emph{score-to-claim inference}: hold source units fixed, compare each claim-induced condition to benchmark-native controls or same-source invariance checks, and apply a declared finite-family support rule. In a fixed set of five open-weight model families, completed records support narrower sentences than endpoint readings alone suggest: supplied-context answer recovery remains separate from support attribution; local supplied-evidence verdict signals leave open-web provenance unmeasured; and SATBench exposes label-surface instability for SAT/UNSAT prediction. The output is a support boundary, a blocked strengthening, and the next measurement or model-selection sensitivity attached to the reported result.

Large language models change their answers when a "verified" source contradicts them. We study this verified-source authority bias across five open-weight families and three closed APIs. On a baseline-correct trivia subset, a verified-source cue endorsing a wrong answer flips 43–87% of responses across seven of the eight models, with a graded hierarchy over cue strengths. We then show this is not generic user sycophancy: across both open-weight families and closed APIs, the same wrong answer endorsed by a verified source vs. by a user produces different behavioral compliance, and a fitted "authority" vector is distinct from a generic assistant/instruction-following direction and from a user-sycophancy direction under causal projection-removal. Applying this vector to neutral prompts induces matched wrong-answer flips, and the same trivia-fit vector transfers without refitting to PIQA (a held-out task) and to multi-turn SYCON dialogues. Projecting the vector out at α=1 reduces wrong-source compliance by tens of percentage points on both trivia and PIQA in four of five open-weight families, while capability checks (MMLU-Pro, GSM8K) stay within noise of baseline at our evaluation sizes.

Large language models can often identify useful semantic structure in formal artifacts, but their outputs cannot be trusted as proofs or solver decisions. We study whether LLMs can instead guide formal reasoning systems through a narrow, solver-verifiable interface. We instantiate this idea in quantified array SMT solving, where modern instantiation-based techniques can fail when key ground terms and equalities are hidden by auxiliary variables, axiom guards, and nested array terms. We observe that many instances are largely driven by a small set of focal ground terms or constraints, similar to the backdoor sets in SAT solving. Building on this, we present LinguaArray, a framework where an LLM proposes a small Semantic Focus Set (ground terms/constraints), while a conventional SMT solver remains the sole proof engine and certifies all results. LinguaArray uses these LLM-suggested, solver-validated finite sets to significantly enhance solver capability, solving additional satisfiable and unsatisfiable instances that the backend alone fails to solve. It interacts within a finite time and then falls back to the backend solver to mitigate potential regressions. Evaluation on 1015 SMT-LIB instances across five logics under 1200s timeout (including LLM latency) shows that LinguaArray solves up to 69.1\% of instances and increases solved-instance counts by 105\%--397\% over the corresponding backends.

Selecting the best large language model (LLM) for a fixed benchmark is often expensive, since exhaustive evaluation requires running every model on every example. Multi-armed bandit (MAB) algorithms can reduce the number of LLM calls by sequentially selecting the next model-example evaluations, thereby avoiding unnecessary computations on clearly underperforming models. Further computational savings can be achieved by predicting the evaluation scores of the various models on the set of examples. In practice, these predictions can be obtained using low-rank (LR) matrix factorization that exploits correlations in the partially observed model–example score matrix. However, such predicted evaluations are not the ground truth: they can be biased and may therefore lead to incorrect identification of the best model. In this work, we propose a principled framework that combines MAB with cheap, predicted evaluation scores without compromising statistical validity. Concretely, leveraging prediction-powered inference (PPI), we derive unbiased estimators for the performance of each model that utilize predicted scores from LR factorization to reduce variance. Importantly, this enables the construction of finite-sample valid confidence intervals in our non-i.i.d. setting, where models are selected adaptively and examples are sampled without replacement. Empirical results on real-world benchmarks demonstrate that our approach reduces the number of required evaluations, leading to meaningful savings in compute and cost, while accurately identifying the best-performing model.

Large language models (LLMs) have made misinformation inexpensive to produce but not to verify, creating a growing asymmetry in the information ecosystem. Under tight time, labor, and budget constraints, media organizations, platforms, and fact-checkers rely on screening to prioritize which content to verify. We introduce **VEX-Bench**, a unified benchmark for evaluating the *verification complexity* of LLM-generated misinformation, as perceived during screening, across models and generation methods. Verification complexity is assessed along multiple dimensions derived from journalistic and fact-checking practices, capturing checkability, harm potential, source credibility signals, imposter legitimacy, and expected verification effort. We define the $\mathrm{VEX}$ score as an integrated measure combining elicitation yield and verification complexity to quantify how generated content consumes limited verification capacity. We construct a benchmark spanning two misinformation categories, 6 high-stakes domains, and 60 real-world topics, and evaluate 7 frontier LLMs and 7 generation methods, yielding 5,880 articles. We employ an LLM-as-judge for scalable evaluation and validate it using content-analysis methodology, including ordinal Krippendorff~$\alpha$ for inter-annotator reliability, complemented by fact-checking agents for verification. Our findings show that no single method dominates all dimensions, underscoring the need for multi-dimensional evaluation. LLMs can generate high-$\mathrm{VEX}$ misinformation at 3$\times$ to 169$\times$ lower cost than agent-based verification. Such content is often prioritized during screening, consuming scarce verification resources and introducing a systematic risk of misallocation in resource-constrained verification systems.

LLM routing matters most in long-horizon applications such as coding agents, deep research systems, and computer-use agents, where a single user request triggers many model calls. Routing each call to the cheapest sufficient model can cut costs without sacrificing quality, yet existing router benchmarks evaluate only which model should answer a whole prompt. They never expose the router-visible prefix at an intermediate agent step, never test whether a cheaper replacement preserves downstream task success, and often rely on online LLM judges at evaluation time. We introduce TwinRouterBench, a step-level routing benchmark with two tracks. The static track provides 970 router-visible prefixes from 520 instances across SWE-bench, BFCL, mtRAG, QMSum, and PinchBench, each paired with an execution-verified target tier estimated under a released downgrade-and-cascade protocol; scoring is deterministic arithmetic over tier labels, trajectory membership, and token costs, with no online evaluator-side LLM judge. The dynamic track supplies a harness that runs routers on the full 500-case SWE-bench Verified suite; in this paper we report a 100-case held-out evaluation disjoint from the static SWE supervision split. At each LLM call the router selects a concrete model from a locked pool, and success is measured by official task resolution and realized API spend. The two tracks support fast offline iteration followed by end-to-end validation under live agent execution. Code and data are available at https://anonymous.4open.science/r/TwinRouterBench-8820.

Language models increasingly act as epistemic proxies, synthesizing evidence from multiple sources to inform decisions. Whether they evaluate the quality of that evidence, or merely aggregate it based on surface presentation, remains poorly understood. We show that models possess the capability to detect fabricated statistics (correct identification rates of $0.76$--$1.00$ for methodology in isolation) but do not recruit this capability during multi-source synthesis, producing similar numeric estimates whether the statistics are fabricated or valid. Specifically, source influence is governed by a methodology-register gate that responds to the distributional register of analytical text but not to numeric validity: for example, statistically impossible confidence intervals receive the same weight as valid ones. The behavioral dissociation replicates across five models from three families (Claude, Qwen, OLMo) and three professional domains. Mechanistic analyses, including causal tracing, linear probes, and component-level attribution, converge on the same account: the model encodes and causally uses a methodology-register representation that transfers across domains (probe AUC $0.83$--$0.92$), while numeric-validity signals, decodable in isolation, are suppressed to chance during multi-source synthesis. Prompting-based mitigations, even an oracle checklist naming the exact statistical checks, produce blanket skepticism rather than selective discernment, and the post-training pipelines we examine reinforce the stylistic shortcut without building numeric verification. Unlike sycophancy, which tracks user preference, this failure tracks whether a source presents as analytically credible, not whether its claims are internally consistent. We term this $\textit{epistemic alignment}$: like preference and safety alignment, the question is not capability but deployment.

Large Language Model (LLM) agents increasingly excel at complex tasks, yet a persistent bottleneck emerges at the boundary of rule-based systems. Here, the core challenge is often not comprehending intent, but routing user intent through a prescribed, executable interface. This paper posits that this routing problem is best addressed not by scaling model reasoning alone, but by treating the interface itself—specifically, Domain-Specific Languages (DSLs) and their attached verifiers—as first-class boundary infrastructure. We argue that an effective DSL functions as a cognitive shortcut: by providing a high-abstraction, strongly constrained representation, it can compress the intent-to-action path, prune invalid actions before they are attempted, and mechanistically verify correctness via its harness. Our central contribution is to frame the deliberate selection, design, and evaluation of such language--harness pairings as an under-studied research variable that directly shapes an agent's effective action space, failure modes, and learning signals. We develop this position through a dual lens of lifecycle and domain analyses, and ground it with targeted case studies.

Adaptive prompt and program search makes LLM evaluation selection-sensitive. Once benchmark items are reused inside tuning, the observed winner's score need not estimate the fresh-data performance of the full tune-then-deploy procedure. We study inference for this procedure-level target under explicit tuning budgets. We propose SIREN, a selection-aware repeated-split reporting protocol that freezes the post-search shortlist, separates splitwise selection from held-out evaluation, and uses an item-level Gaussian multiplier bootstrap for uncertainty quantification. In a fixed-shortlist regime with smooth stabilized selection, the estimator admits a first-order item-level representation, and the bootstrap yields valid simultaneous inference on a finite budget grid. This supports confidence intervals for procedure-performance curves and pre-specified equal-budget and cross-budget comparisons. Controlled simulations and MMLU-Pro tuning experiments show that winner-based reporting can be optimistic and can change deployment conclusions, while SIREN remains close to the finite-sample reporting target.

Larger language models solve harder problems, but do they also produce proportionally better explanations? We study in-context explanation transfer across 40 open-weight models (360M--72B) and 5 receivers (360M--7B), using problems drawn from 13 benchmark sources and evaluated with domain-specific metrics. Our central finding is an \emph{articulability ceiling}: under equal token budgets, mid-range models are better teachers than frontier models. Transfer peaks at 170--220 words and then drops sharply beyond 300 words, falling below even the shortest explanations. Large models partly compensate through verbosity: longer explanations yield higher total transfer on average, but each additional word carries less pedagogical value, so a 360M model delivers more transfer value per word than a 72B model. We also find that RLHF improves pedagogical clarity: instruct models score better on LLM-judge evaluation than base models despite lower lexical overlap, and teacher correctness dominates all other factors. Code and logs are released with the paper.

LLM judges are increasingly used to rank systems and validate claimed improvements, but their validation often reduces to a single question: how well do judge scores correlate with aggregated human ratings? We argue that this question is too fragile. A judge can correlate with human averages while collapsing the rating scale, compressing meaningful margins, or confidently separating systems that humans do not distinguish. Furthermore, a correlation-based workflow always returns a leaderboard, even when the human benchmark cannot support one. We propose a benchmark-conditional measurement framework for LLM-judge validation. Stage 1 audits the human benchmark as a reference ruler using a many-facet ordered Rasch model, estimating rater/item fit, conditional precision, and supported pairwise resolution. Stage 2 calibrates each judge as a linked measurement instrument with judge-specific ordinal thresholds. Stage 3 evaluates out-of-sample system-level fidelity using normalized signed-margin fidelity, false-separation diagnostics, empirical transfer intervals, and distributional EMD. On SummEval and HiTZ-BASSE, the framework yields three outcomes: selection, warning, and refusal, depending on the ruler and judge pool. Calibrated predictions improve observable human-score distribution EMD in 57/81 judge–dimension cases, while the remaining failures are concentrated in the regions flagged by the ruler or judge-instrument audits. Together, these results instantiate benchmark-conditional LLM-judge validation as a protocol for deciding when judge-based claims are warranted, uncertain, or unsupported, addressing a central blind spot of correlation-based judge evaluation.

Evaluating enterprise LLMs on file-grounded Question Answering (QA) requires large, labeled corpora of realistic enterprise PDFs, yet real documents are sensitive, scarce, and impossible to annotate at scale. Existing synthetic approaches produce visually crude renders, annotate ground truth post-hoc (introducing hallucination risk), or are non-reproducible one-off scripts with no diversity control. We present SynthDocQA, an Intermediate Representation(IR)-driven pipeline that co-generates photorealistic multi-page enterprise PDFs and provenance-grounded evaluation pairs in a single pass. A compact Document Generation IR encodes topic, content schedule, and perturbation profile before rendering begins. Templates are retrieved per page via Contrastive Language-Image Pretraining (CLIP) embedding-based cross-modal search over a 10,000-layout library; a 15-stage physical degradation pipeline models five scan quality tiers from pristine to poor. Every QA assertion is derived analytically from the same structured artifact that populated the page, making hallucination structurally impossible, and a hierarchical fork-based Random Number Generator (RNG) guarantees exact reproducibility from a single integer seed. A reference 100-document corpus spans 50 enterprise domains, 4,374 pages, 16,589 QA pairs, and 21,402 assertions at zero annotation cost.

Large language models confidently produce outdated answers, and no existing method can detect them. We show this is not an engineering failure but a structural one: temporal drift, whether a stored fact has changed since training, is encoded as a direction in the residual stream geometrically orthogonal to both correctness and uncertainty. Any method operating on correctness or uncertainty signals is therefore blind to drift by construction. We verify this across six instruction-tuned models. A linear probe trained directly on drift labels achieves AUROC $0.83$--$0.95$; methods based on token entropy, semantic entropy, CCS, and SAPLMA all remain near chance ($0.49$--$0.57$). Five tests confirm the geometric orthogonality: weight cosines ($|\cos| \leq 0.14$), score correlations ($|r| \leq 0.20$), bidirectional null-space projection ($|\Delta| \leq 0.008$), iterative null-space projection with $k{=}10$, and difference-of-means dissociation. Mechanistically, the MLP retrieval circuit produces identical dynamics for stale recall and confabulation ($r > 0.81$, six models), explaining why output confidence cannot separate them. A cross-cutoff experiment holds inputs constant and varies only the model: the probe fires on the model whose training predates the fact's transition and stays silent otherwise ($P(A{>}B) = 0.975$--$0.998$, twelve model pairs), confirming it reads model-internal knowledge state rather than input properties. Our code and datasets will be publicly released.

We introduce TextSeal, a state-of-the-art watermark for large language models. Building on Gumbel-max sampling, TextSeal introduces dual-key generation to restore output diversity, along with entropy-weighted scoring and multi-region localization for improved robustness. It supports serving optimizations such as speculative decoding and multi-token prediction, and does not add any inference overhead. TextSeal strictly dominates baselines like SynthID-text in detection strength and is robust to dilution, maintaining confident localized detection even in heavily mixed human/AI documents. The scheme is theoretically distortion-free, and evaluation across reasoning benchmarks confirms that it preserves downstream performance; while a multilingual human evaluation (6000 A/B comparisons, 5 languages) shows no perceptible quality difference. Beyond its use for provenance detection, TextSeal is also "radioactive": its watermark signal transfers through model distillation, enabling detection of unauthorized use.

Knowledge distillation generally assumes a strong-to-weak relationship where stronger teachers yield better students. In this work, we examine this assumption about distillation in large language model (LLM) pretraining. By varying architecture sizes and training token budgets, we create strong-to-weak, same-level, and weak-to-strong teacher-student relationships. We study distillation's effectiveness under these relationships with different mixes between language and distillation loss. Three findings emerge: (1) with proper loss mixing, weak-to-strong and same-level distillation improves over standard pretraining, where even small and undertrained teachers benefit large students; (2) making the teacher stronger can lead to saturated or even reversed gains; (3) distillation improves generalization (out-of-domain, downstream) more readily than in-domain fitting. Our results provide practical guidance for choosing the teacher model with the proper loss for more effective LLM pretraining distillation.

Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve the reasoning capabilities of large language models, but its reward is derived only from the correctness of the final answer and provides no supervision over intermediate reasoning steps. Recent rubric-based methods, such as Rubrics as Rewards (RaR), introduce finer-grained supervision by scoring rollouts against a structured set of evaluation criteria. However, the resulting rubric scores are still aggregated into a single scalar that is applied to the entire response, leading to three structural weaknesses, namely loss of the multi-criterion rubric structure, uniform supervision of correct and incorrect reasoning steps, and reward hacking in the trained model through unbounded self-correction. On a sample of 1{,}000 problems, we find that 18.2\% of steps within answer-correct responses are themselves wrong yet positively rewarded, while 49.9\% of steps within answer-incorrect responses are in fact correct yet penalized. Therefore, we introduce \textbf{Step-wise Rubrics as Rewards (SRaR)}, an RLVR framework that (i) uses an LLM judge to attribute each rubric item to a specific reasoning step, (ii) normalizes the per-step rubric scores across rollouts so that only steps whose quality varies produce a learning signal, and (iii) combines the resulting per-step reward with the standard outcome reward through a decoupled advantage estimator that keeps the outcome-driven baseline stable. To support training, we further build a 16K-problem rubric dataset by contrastively distilling rubric items from correct and flawed reasoning paths sampled from a strong model and verified against the ground-truth answer. Across six mathematical reasoning benchmarks spanning multiple difficulty levels, SRaR improves the average accuracy over RaR by \textbf{3.57} points on Qwen3-8B-Non-Thinking and by \textbf{2.75} points on Qwen3-32B-Non-Thinking, raises the Faithful Reasoning Rate on AIME~2025 from 34.5\% to 46.7\% (every correct-answer trajectory of SRaR uses entirely correct reasoning steps), and reduces the rate of self-correction looping from 48.1\% to 26.5\%.