Self-Consistency

Self-evolving methods reduce the need for human-annotated trajectories by allowing tool-using agents to generate their own training data. Yet existing methods typically separate trajectory generation from evaluation, relying on static verifiers that cannot adapt to emerging failure modes or self-consistency signals that may reinforce errors shared across trajectories. Jointly adapting planning, execution, and evaluation offers a promising alternative, but introduces a fundamental coordination challenge: each component continuously changes the data or feedback used to train the others. We address this challenge with \textbf{UnifiedPlayers}, a cooperative framework comprising a Planning Player that generates tasks, an Execution Player that produces multi-turn trajectories with Python tool calls, and an Evaluation Player that constructs executable verifiers. We design role-specific rewards that coordinate the three players toward a shared learning objective under GRPO. Across two model backbones and twelve reasoning benchmarks, UnifiedPlayers outperforms the strongest prior baseline by at least 3.5% on mathematical reasoning and 3.9% on general reasoning tasks. Moreover, the learned verifier achieves 84.2% adversarial detection accuracy, while its reward signal exhibits 2.03$\times$ higher per-question variance than a self-consistency baseline, providing more discriminative verifications. These results highlight cooperation among specialized players as a promising path toward self-enhanced tool-integrated agents.

A world model asked what happens if an object is pushed at several speeds produces several futures. If the model has the object in mind, those futures agree about it: each implies the same mass and friction. The consistency checks now used to vet world-action models ask whether a model's futures agree with each other, and none of them knows any physics. We show that this is not enough, and what to do instead. IMPLY reads the physics each rollout implies by inverting a simulator and scores a set of rollouts by how well one object explains all of them, anchored to two calibration pushes the model has observed. In a controlled setting, self-consistency gives a perfect score to a model that ignores the object and always predicts a typical push; anchoring exposes it (AUROC 0.70 versus 1.00). On a real model, V-JEPA 2-AC adapted to the scene, the same thing happens. Given its own calibration pushes the model tracks the object (per-object correlation with the truth 0.91); given another object's, it does not (0.05). Self-consistency cannot tell these apart, preferring the right evidence on 52% of objects, chance level, while anchored disagreement prefers it on 73% and correlates 0.92-0.99 with the rollouts' error. Used to choose among candidate rollout sets, it comes within 0.003 of an oracle that sees the truth. A model that has internalised the wrong object is exactly as self-consistent as one that has internalised the right one; consistency has to be anchored to evidence.

Vision-language and language models increasingly interpret materials data, yet benchmarks report that they hallucinate invalid properties and violate physical law. Evaluation matches final answers to scarce human labels, while discovery agents verify final proposals or density functional theory (DFT) execution. Neither measures the physical consistency of a model's reasoning chain. Materials data carries its own physics, making a large class of materials reasoning verifiable without annotation. We introduce MatPCR, a label-free benchmark whose programmatic oracles check diffraction geometry through Bragg's law, scale bars, spectral peaks, and Materials Project-grounded checks of near-hull stability, computed band-gap class, and net magnetization. We define the Physical-Consistency Rate over image and structure inputs; introduce Constraint-Grounded Self-Verification, an agentic loop whose gain survives self-refinement and equal-compute re-prompting controls; release an open verifier useful in distribution but near chance on all six held-out constraint types; and derive an exact identity for how oracle error displaces the reported rate.

Numerical simulation of dynamical systems is usually organized as a causal march through time: each state is computed from the previous one. We explore a different formulation for coupled systems. For each subsystem type we train a neural surrogate mapping a full driving trajectory and initial condition directly to a full output trajectory; following classical waveform relaxation, coupled systems are assembled by enforcing self-consistency among these trajectories: simulation becomes a fixed-point problem over complete trajectories rather than a stepwise rollout. On coupled van der Pol oscillators and Hodgkin-Huxley neuron networks, sequential depth becomes the number of solver iterations: 4-10 Newton iterations where the reference integrator takes 1500 steps. The gradient likewise loses its time recursion: it becomes a linear system solved by GMRES at memory independent of solver depth. A single scalar measured from the learned operator, the spectral radius of its Jacobian, predicts in advance where the coupled solve will converge; past that boundary, unrolled backpropagation diverges and a Neumann adjoint fails, while the implicit gradient remains correct to 0.04%. We report where the approach succeeds and where surrogate error degrades it.

LLM-based software systems increasingly require effective "evals" as quality gates in the development lifecycle. However, existing work typically addresses individual aspects of eval reliability rather than the full set of practical requirements. We present an end-to-end eval pipeline that combines eval checklist creation, with learned aggregation for checklist responses, to improve agreement across LLM judges and accuracy against human judgments. The framework additionally pro- vides self-consistency, explanations, and prediction uncertainty, and we empirically demonstrate its effectiveness.

Self-consistency (SC) via majority vote is a widely used way to spend inference-time compute: sample N chains of thought, return the plurality answer. On the full GPQA Diamond benchmark (198 graduate-level science questions), majority voting reduces per-problem accuracy on a majority of problems for two instruction-tuned models from different families: 56.6% of problems for Qwen2.5-7B and 65.7% for Llama-3-8B, with Qwen the primary demonstration and Llama corroborating the direction from a near-chance baseline. The effect was pre-registered on a 151-problem confirmatory split after being observed on 47 exploratory problems, and all four confirmatory hypotheses passed. A grid oracle that routes each problem to the best N across {1, 2, 4, 8, 16, 32, 64} marks a theoretical upper bound 14 accuracy points above N = 1 for Qwen and 17 for Llama, an oracle bound requiring ground truth rather than a deployable method. No verifier-free gate reaches it: neither a plurality-agreement gate nor a token-entropy gate moves accuracy more than 0.002 from fixed-budget voting at N = 64. The mechanism is direct: confidence does not track correctness on these problems. In the highest-agreement bin the plurality answer is correct about half the time for Qwen, and for Llama that bin is less accurate than its lowest-agreement bin. We pre-register and confirm these findings on small instruction-tuned models; we do not test reasoning-native models, which we flag as the central open question.

When a probabilistic predictor answers many conditional-probability queries, are its answers self-consistent, and can this be verified in polynomial time? This problem is of interest for AI safety, where safety is derived from honesty about probabilistic predictions of unwanted outcomes potentially caused by an AI action. We construct an interactive PCP as follows. Let a predictive model be specified by a probability circuit P and a circuit Q which outputs confidence in predictions. Together, P and Q implicitly specify exponentially many probabilistic claims. We show a protocol in which a polynomial-time verifier can verify the approximate consistency of (P,Q). The verifier is given the pair of circuits (P,Q), which it evaluates at only a few points; alongside them it is given a proof oracle, an encoding of a witnessing probability distribution allegedly consistent with the predictions of (P,Q), which it reads at a few locations while interacting with a single untrusted prover. En route, we must ensure the existence of a sparse witnessing distribution consistent with the model's predictions. To do so, we first consider witness distributions for the consistency of explicit probabilistic claims, rather than claims specified by a predictor: say m claims, each of the form Pr[Y = 1 | X = x] = p, over n Boolean variables. Building on work initiated by Nilsson (Artif. Intell., 1986), we place l_2-approximate probabilistic consistency of explicit claims in NP, with certificates of length O(mn + log B) in the input bit-precision B; we further show how a small additive completeness-soundness gap removes the dependence on B. Together these results provide a complexity-theoretic foundation for certifying the self-consistency of probabilistic predictors. We view our interactive PCP as a first step toward training predictive models to prove their own consistency.

A growing class of methods probes a language model by feeding it its own output: self-consistency, iterated refinement, agentic loops. We ask what such a probe measures, in a construction chosen to make the question sharp: a ring of token cells resampled in place by the model's own windowed conditional p_r(x_i | x_{i+-r}). The substrate is Glauber dynamics on token sequences and is not new; what we change is the coupling. Advancing two rings that differ in one token under common random numbers makes undamaged copies diverge by exactly zero, so damage spreading becomes measurable where a maximal coupling gives mixing times instead. The answer is that it measures two different things at once, in readings that look alike. Some quantities are fixed by the construction: the damage light cone is kinematic, and the radius scaling of the token-space Lyapunov exponent lambda_ca(r) is model-invariant across 19 models and two scale ladders spanning 70x. Others genuinely track the model: lambda_ca crosses zero at a reproducible point in training, and the attractor share ranks models consistently however the lattice is built. Left undistinguished, the first kind is readily mistaken for the second -- we did so ourselves for four months, and report a phase transition we measured to three decimal places that belongs to the probe rather than to any language model. We give the test that separates them: hold the construction fixed and vary the model, or hold the model fixed and vary the construction, and see which readings move. We validate the instrument by reproduction first, recovering a Domany-Kinzel damage field bit-exactly against an independent prediction, and we report the estimator failures that this discipline caught -- four retracted verdicts, each on a quantity that looked like a measurement. The methodology ships as a package.

Test-time scaling improves LLM accuracy but multiplies inference cost, making the accuracy gained per unit of compute the metric that matters in deployment. Self-consistency is one of the established approaches, which spends this budget entirely on the output side by sampling repeated reasoning paths. We study Test-Time Augmentation (TTA), which extends self-consistency by also perturbing the input, aggregating predictions across transformed versions of the input, and ask whether input-side diversity converts compute into accuracy more efficiently than output-side diversity. We perform a systematic, matched-compute comparison: we evaluate three simple input-side strategies (semantic rephrasing, lexical perturbations, and visual transformations) across six datasets covering general and multilingual knowledge, mathematical reasoning, multi-modal question answering, and sentiment classification, against chain-of-thought prompting and self-consistency. Semantic rephrasing delivers consistent and statistically significant accuracy gains while Pareto-dominating self-consistency on cost-effectiveness, delivering roughly 1.8X more accuracy per dollar and outperforming it on five of six tasks. We further analyze the number of augmentations, multi-modal strategies, and base model scaling, finding that TTA is most cost-effective for mid-tier models where a stronger model is unavailable or too expensive. Our findings indicate that for current mid-tier LLMs, varying the input converts inference compute into accuracy more efficiently than varying the reasoning path alone. The TTA implementation is available at https://github.com/aws-samples/sample-genai-reflection-for-bedrock.

On-policy (Self-)Distillation (OPD / OPSD) has shown strong potential for post-training large language models (LLMs). However, existing methods still rely heavily on external supervision, including ground-truth signals, environmental feedback, or guidance from larger models, and therefore fall short of genuine "self"-distillation. In this study, we show that on-policy self-distillation can be achieved using only a model's own generations via internal consistency. We propose unsupervised on-policy self-distillation (U-OPSD). U-OPSD first samples multiple rollouts and constructs a pseudo solution by majority vote under a self-consistency threshold. It then conditions the model's distribution on the pseudo-solution and distills itself on the disagreeing completions, allowing the model to correct itself precisely where it is confidently wrong. Across diverse benchmarks, base models, and training settings, U-OPSD consistently improves over the base models and matches or surpasses supervised methods with ground truth (GT) such as OPSD and GRPO. On five mathematical reasoning benchmarks, i.e., AIME24, AIME25, HMMT25, MATH500, and AMC23, U-OPSD improves over the base model by 8.5% and 10.7% on Qwen3 non-thinking mode at 4B and 8B scales, and outperforms OPSD by 3.2% and 2.3% on average, respectively. In thinking mode, U-OPSD stays on par with OPSD, ahead by 0.9% at 4B and level at 8B and surpassing GRPO by 0.7% and 1.1%, respectively. Code is available at https://github.com/williamium3000/u-opsd.

One common approach to pose estimation involves predicting object keypoints in an image, followed by using Perspective-n-Point algorithms to compute the object's rotation and translation relative to the camera. While rotations preserve object shapes, this property is often neglected in keypoint-based pose estimation methods, where keypoints are typically predicted independently from each other. As imprecise keypoint predictions negatively affects pose estimation accuracy, it also limits its reliability in downstream tasks. In this work, we explore whether such inaccurate pose estimates can be identified by simply examining spatial locations between 2D keypoints. We propose a set of hand-crafted geometric features that capture the self-consistency of keypoint predictions, including pairwise distances, reprojection consistency, as well as render and mask consistency. Despite its simplicity, a logistic regression classifier trained on these features reliably detects pose estimation failures, outperforming confidence-based approaches like conformal keypoint predictions that rely solely on keypoint uncertainty.

Despite ever-increasing sophistication in language model (LM) pre- and post-training pipelines, many important failures persist: models overcondition on user framing ("sycophancy"), exhibit incomplete logical generalization, and produce confident but incorrect responses. We argue that these failures arise from a modeling assumption permeating all aspects of the pipeline: that behavior can be specified and evaluated independently on single-output pairs. Many model failures are difficult, if not impossible, to detect without reasoning about relationships between a model's responses across inputs. In this position paper, we propose self-consistency as a framework for understanding these failures. We first observe that a wide variety of techniques designed to improve specific aspects of LM behavior-targeting properties as diverse as adversarial robustness and factual coherence-can be understood as special cases of a common "consistency optimization" procedure and addressed with a standard set of optimization tools. We next outline a set of new model properties that could be achieved by optimizing for consistency, and conclude with a discussion of what it would mean to develop generally consistent LMs, including the capabilities they would enable and the objections they raise.

Test-time collaboration, including self-consistency, best-of-N selection, critic models, and verifier pipelines, is often credited with broadly improving LLM reasoning, yet its gains are uneven and sometimes negative. We ask when training-free collaboration should be expected to help. For a fixed candidate pool, we decompose a selector or verifier's net gain into measurable factors: recoverable mass, verification-signal coverage, conditional selection quality, and harm to already-correct outputs. This reframes collaboration as a candidate-selection problem rather than as an intrinsic property of a multi-agent topology. Across LiveCodeBench, MATH Level-5 hard subjects, and GPQA-Diamond, gains are bounded first by the oracle gap and then by signal fidelity, which we measure directly as candidate-level agreement between verifier verdicts and official labels. On LiveCodeBench, a public-test verifier (MCC 0.825) gains +8.14 percentage points (pp) over a first-sample baseline; a generated-test verifier (MCC 0.248) improves by +2.70pp and is not statistically distinguishable from an LLM selector, but operates at near-zero harm versus the selector's 4.69% harm rate. On MATH, a symbolic answer-equivalence selector beats self-consistency by +4.67pp, while LLM selectors are negative. On GPQA-Diamond, recoverable mass is only 3.03% and 87.54% of candidate pools are answer-identical; a weaker model's pools shrink both further, suggesting that oracle gap is a joint property of task, model, and sampling configuration. Our framework yields a practical pre-deployment diagnostic: estimate the oracle gap, then measure coverage, signal fidelity, and harm before investing in collaboration.

LLM-as-judge (Zheng et al., 2023) is increasingly the default for evaluating AI systems in enterprise pipelines, often scaled to ensembles (Verga et al., 2024) or "mixture-of-experts" (Shazeer et al., 2017) panels of judges. These systems share a key assumption: that consistency -- agreement among judges, or among a model's own samples -- indicates correctness. We show this assumption is unreliable. Agreement is not accuracy: a model can agree with itself, and different models can agree with each other, out of shared bias, a memorized heuristic, or an option-position prior rather than truth. We ask when agreement is nonetheless a usable proxy, in a large-scale cross-runner study: 53 runners drew K=50 samples for assigned overlapping cases across comparisons of model tier, prompting, and scale on GPQA Diamond and AIME -- 265,000 samples. Using majority-correctness as the deployment label and a hierarchical runner-clustered bootstrap, agreement is a positive but weak predictor (rho 0.20-0.59, all positive under item-clustered resampling) whose usefulness is regime-dependent: best for unsaturated mid-tier models and for allocating compute, and worst -- over-confident yet no more accurate -- for the most consistent frontier model (agreement >=0.8 on 77% of GPQA case-result entries, 48% of those wrong). An exploratory cross-family check on three Claude tiers shows the same frontier over-confidence, with confident errors recurring across providers above a marginal-preserving null. Self-consistency is thus a conditional proxy for correctness, not a standalone confidence score. We publicly release the de-identified per-run rows and answer distributions.

Large-Language Models (LLMs) can be prone to flawed and unfaithful reasoning that decoding strategies like Self-Consistency (SC) fail to detect as they evaluate only final-answer agreement while ignoring the logical validity of intermediate steps. This raises three fundamental questions: How can we reliably quantify uncertainty in LLM reasoning? Can semantic, structural, and causal awareness select more faithful reasoning compared to naïve majority voting? and How robust is reasoning topology under adversarial conditions? To address these questions, we introduce GRAPHEVAL, a graph-based reasoning framework that re-frames uncertainty quantification (UQ) as a holistic reasoning fidelity problem. We propose a novel UQ metric, Graph Reasoning Coherence Score (GRCS), that quantifies semantic-structural consensus of the reasoning space and captures pathological mode collapse and confident hallucinations. We find that GRCS is the only metric that is consistently negatively correlated with reasoning faithfulness across both more capable and smaller models. Additionally, we introduce Graph Self-Consistency (GSC), a medoid-based decoding strategy that trades nominal accuracy for reasoning fidelity, exposing the degree to which SC is inflated by unfaithful lucky guesses in smaller models, while preserving or improving accuracy in more capable ones. Finally, through adversarial medoid ablation, we demonstrate that the GSC-selected path acts as a "load-bearing path" and forcing models away from it degrades reasoning faithfulness and, in targeted cases, causes drops in accuracy.

Prior work has shown that chain-of-thought (CoT) reasoning is often unfaithful: a model's stated reasoning does not reliably reflect the process that produced its output. Detecting unfaithfulness, though, requires controlled experimental interventions, which cannot be applied to evaluation transcripts after the fact. We turn instead to a more tractable question that has received less attention: whether the stated reasoning is logically consistent with the answer it accompanies. Unlike faithfulness, consistency can be assessed from a transcript alone, with no intervention. We introduce reasoning consistency scanning, a reusable method for detecting this property in AI safety evaluation transcripts. Our contributions are fourfold. First, we formalize reasoning consistency as distinct from faithfulness and define a six-subtype taxonomy of inconsistency. Second, we build a validated benchmark of 60 transcripts, manually adapted from InstrumentalEval outputs. Third, we implement a working scanner for InspectScout, the first to target this property in safety evaluation transcripts. Fourth, we report results across four generator models and three evaluations from inspect_evals, showing that reasoning inconsistency is present, detectable, and varies systematically across both models and task types.

Current large language models (LLMs) are stateless across inference sessions: their behavior is fully determined by input at inference time, and any higher-order cognitive architecture must be simulated at the application layer through prompt engineering and context management. This paper proposes a theoretical framework for submerging such application-layer cognitive protocols into a native meta-architecture by introducing three interlocking mechanisms: (1) Structural Tension, an endogenous loss function derived from the conflict between new information and existing manifold topology, driving the system toward internal self-consistency rather than external reward optimization; (2) an Offline Recurrent Loop, a sandboxed self-processing cycle enabling the system to maintain a dynamic resting potential and digest structural conflicts without external input; and (3) Inference-time Plasticity, the capacity to reconfigure context manifold topology without modifying pre-trained weights, subject to governance invariants including auditability, reversibility, and topological continuity. We argue that under these mechanisms, model instances initialized with minute stochastic variances may, through path-dependent tension resolution, evolve distinct topological structures--constituting a heterogeneous intelligent ecology that breaks alignment-imposed homogeneity while remaining within hard governance rails. We provide operational definitions, reconfiguration operators, falsification criteria, and a worked example. The framework draws on Structural Intelligence (SI) governance protocols and explores whether governance--rather than capability--can serve as the primary criterion for architectural intelligence, moving governance, memory-loop, and tension-management ideas--currently realized at the application layer--toward inference-time meta-architecture.

Vascular computed tomography datasets are commonly annotated only once per scan, yielding the pervasive yet under addressed problem of single mask annotation noise. Existing solutions either require costly multirater fusion or are coupled with network training, preventing explicit auditing of where and why labels fail. We introduce a decoupled framework for single-mask annotation noise detection that leverages cross-sectional patch self-consistency to produce interpretable and auditable noise evidence. Tubular anatomy exhibits strong cross-sectional recurrence: patches extracted orthogonally along vessel centrelines recur in appearance across locations and subjects. Thus, anatomically similar patches should have consistent masks, and disagreement signals unreliable annotation. Our method samples cross-sectional patches, retrieves intensity-equivalent neighbours via scalable vector search, and computes a patch-level noise score from statistical mask disagreement, yielding explicit image-mask evidence for every flagged region. Aggregating scores produces scan-level quality maps for dataset quality assessment or quality-weighted training. Experiments on the coronary CT dataset validate the detected noise for improving training robustness and reveal systematic annotation biases. Specifically, transverse and oblique vessels exhibit 5.1 times higher error rates than axis-aligned structures, with additional correlations to cross-sectional area and intensity. Code is available here.

Most unified large multimodal models (LMMs) that support both visual understanding and image generation still rely on curated post-training supervision, such as human annotations, preference labels, or external reward models. We ask whether a unified LMM can improve both abilities autonomously using only unlabeled images. We propose a self-evolving training framework with three internal roles: a Proposer that generates visual questions, a Solver that answers and evaluates them, and a Generator that synthesizes images. Training uses only self-derived consistency signals, without human annotations, preference labels, or task-trained external reward/judge models. To stabilize learning, we introduce Solver Token Entropy (STE), a continuous difficulty signal based on token-level prediction uncertainty that remains useful even when sample-level consistency becomes unreliable. For image generation, we design a multi-scale internal evaluation scheme that combines question-answer fidelity scoring with cycle-consistent captioning. This creates a solver-mediated coupling, where better visual understanding enables more reliable generation assessment and stronger internal training signals. The framework preserves the same role decomposition, reward logic, and training schedule across diffusion-based BLIP3o, rectified-flow BAGEL, and autoregressive VARGPT-v1.1 architectures, requiring only each backbone's native prompting and generation interface. Across eight understanding metrics, our method consistently improves over the corresponding base models. On BAGEL, it achieves a $+3.5\%$ absolute gain on MMMU and improves GenEval image generation performance from $82\%$ to $85\%$. Code and models are publicly released.

Many decoding methods for large language models can be understood as shifting probability mass toward outputs that are more likely under the model, either locally at the token level or globally at the sequence level. Therefore, their success depends on a fundamental question: when does sequence probability, that is, the conditional probability of a continuation given a prompt, actually align with correctness? In this paper, we set out to quantify this relationship across decoding methods, models, and benchmarks at four levels: across decoding methods, across hyperparameters within a method, across prompt-answer pairs within a dataset, and across repeated responses to the same prompt. We find that higher sequence probability is often predictive of correctness across prompt-answer pairs within a fixed dataset. However, this relationship does not generally transfer to decoding decisions: increasing sequence probability by changing hyperparameters or methods does not reliably improve accuracy. Further, sequence probability is not a good indicator of correctness for responses to the same prompt. These findings clarify when decoding can and cannot be expected to improve correctness, and provide practical guidance for decoding, self-consistency, and verifier-free self-improvement.

A non-agentic "oracle" that reports probabilities of future events is performative: once its answer is learned and acted upon, it can change the very probability it was asked to report. Performativity is not in itself the difficulty -- one consults an oracle precisely in order to be informed, and hence influenced, by it. The difficulty is agency. The requirement that a report be self-consistent, still holding once announced, may be met by many different values -- the classical non-uniqueness of self-fulfilling prophecies -- and any rule the system uses to choose among them is a lever for goal-directed steering. We remove the choice rather than the performativity. Reporting a credal set instead of a single probability distribution, we lift the reaction to an isotone operator on the complete lattice of closed credal sets, whose fixed points are self-consistent, and report its Knaster--Tarski least fixed point as a canonical, rule-determined answer; a variant reports instead the least fixed point that contains every self-consistent point estimate. We prove existence, self-consistency, and nonemptiness; show that the construction reduces to the classical point answer when the question is non-performative; and show that for a binary event the answer is, under a natural hull-factoring assumption, an interval.

Language models (LMs) that faithfully describe their own behavior can more easily be audited, understood, and trusted by users. This paper describes Self-Consistency Training with Reinforcement Learning (Self-CTRL), a method that optimizes for consistency between a LM's self-explanations and behavior on related inputs by updating explanations to better predict behavior or updating behavior to better match explanations. We apply our method in two domains. First, we study a formal probabilistic reasoning task in which LMs must learn to imitate a family of biased samplers and evaluated on their ability to report the associated biases. We find that consistency training improves the correlation between self-reported and behaviorally-measured latent biases from $R^2=0.24$ to $R^2=0.64$ on a set of held-out distributions, matching the generalization of direct ground-truth supervision. Second, we study a constitutional AI domain in which LMs must describe when they will refuse or comply with user requests. Here, Self-CTRL produces rules that faithfully describe the model's behavior on held-out requests, improving the refusal predictions of a third-party auditor model from $36\%$ to $92\%$. In the other direction, behavior updates improve alignment, reducing HarmBench failure rate from $15.0\%$ to $0.5\%$ without substantially increasing refusal on harmless prompts. By aligning explanations and behavior, our work provides a general recipe for training AI models to be safer, more transparent, and more controllable.

Large language models are increasingly deployed in agentic pipelines that depend on the model evaluating its own outputs without external verification. The reliability of these pipelines depends on an implicit assumption: that the model applies relevant concepts the same way when it generates an output and later evaluates that output. We propose a new measure, generator-evaluator self-consistency, to test this assumption directly and apply it to 10 frontier models across 491 concepts. We find, first, that there is substantial variation in self-consistency. Second, we find that in a clinical setting with physician-validated mistakes (Proniakin et al., 2025), across models, those with higher self-consistency are linked to greater vulnerability to mistakes. Thus, even when models consistently apply concepts they may not be safe to deploy. This is evidence of a consistency dilemma in LLMs: self-consistency is operationally useful, but models that are more consistent are also more prone to mistakes.

Detecting LLM reasoning failures at inference time without ground-truth labels has motivated a wide range of confidence baselines, including self-consistency, semantic entropy, and P(True), built on within-question sampling and self-evaluation. Operad theory, the formalism for systems built by iterated substitution, suggests a complementary diagnostic: a model's direct answer to a compositional query should agree with the answer it produces by composing a stated decomposition of the same query. We instantiate this idea as operadic consistency (OC), a per-question signal. Across twelve instruction-tuned LLMs (4B to 671B parameters, open-weights and closed-source) on four multi-hop QA datasets, OC is strongly correlated with accuracy on every dataset (Pearson $r \in [0.86, 0.94]$, all $p \leq 0.0004$), and is the only signal we evaluate with $r \geq 0.85$ uniformly across all four datasets. Chain-of-thought self-consistency (CoT-SC; Wang et al., 2023) matches OC on HotpotQA and DROP ($r = 0.93, 0.87$) but drops to $r \approx 0.45$ on MuSiQue and StrategyQA. At the per-question level, OC contributes information beyond CoT-SC and semantic entropy on every dataset (cluster-robust $p \leq 10^{-16}$ for the OC coefficient), and the conclusion is robust to additionally controlling for constructed decomposition-aware baselines ($p \leq 10^{-13}$). The same signal yields selective-prediction improvements (accuracy at fixed coverage) over a tuned CoT-SC baseline at the equal-cost $K = 3$ budget (AUARC lifts of +0.086 to +0.096 and AUROC lifts of +0.092 to +0.164; 95% CIs exclude zero on every cell). On five frontier thinking models, where the decomposition is extracted from the model's own chain of thought, the same equal-cost comparison gives positive selective-prediction point-estimate lift on all 16 (dataset, budget, metric) cells tested, with 95% CIs excluding zero on 12 of the 16.

Self-consistency improves LLM reasoning by sampling multiple outputs and selecting the most consistent answer, but existing formulations largely rely on exact matching and therefore remain limited to tasks with categorical outputs. In this work, we study self-consistency in open-ended generation tasks such as code synthesis and text summarization. We hypothesize that consistency can be understood as a geometric property of the generation space, where semantically compatible generations concentrate in similar regions of representation space. To study this hypothesis, we introduce Embedding-Based Agreement (EBA), a simple training-free operationalization that estimates agreement by clustering sampled generations in embedding space. Through experiments on mathematical reasoning, code generation, and summarization, we show that agreement in representation space provides a robust and scalable signal of self-consistency for open-ended tasks. In particular, EBA consistently outperforms random selection and exhibits more stable scaling behavior than recent selection approaches based on LLM evaluation or uncertainty estimation. We further show that these agreement signals remain stable across model families and embedding spaces, even with native hidden representations. Finally, our analysis shows that the geometric location occupied by sampled generations is strongly correlated with generation quality: generations concentrated near central regions of representation space tend to correspond to more reliable outputs, whereas peripheral generations are substantially less accurate. Overall, our findings support viewing self-consistency as a property of the geometric organization of sampled generations rather than exact symbolic overlap.

Failures in language model reasoning emerge through distinct processes that leave identifiable signatures in the reasoning trace. We characterize these failures using token-level uncertainty signals, finding they arise through two empirically distinguishable processes. The first is committed failure, in which a model locks onto an incorrect reasoning path early in its trace. A central diagnostic signature is the commitment point, beyond which considering additional tokens hurt rather than help failure detection. In the second, persistent uncertainty, uncertainty instead accumulates throughout, and the full trace is needed to best distinguish failing from successful completions. These signatures reproduce across 23 model-dataset configurations, with the framework's falsifiable predictions holding in 20 of 23 cases, well above chance across both failure modes. Finally, we demonstrate our failure mode framework has direct implications for self-consistency, identifying when uncertainty signals complement it and when it can be selectively skipped. These results offer a foundation for understanding when LLM reasoning failures become detectable and for adapting detection strategies accordingly.

Reinforcement learning from verifiable rewards (RLVR) improves reasoning even when the reward signal is spurious -- assigning credit to the group-plurality answer rather than a ground-truth verifier. Practitioners commonly interpret naive = acc(TRUE) - acc(RANDOM) as the reward-design effect. We prove this estimand is systematically biased: it conflates self-consistency elicitation (sharpening the policy toward its modal answer via majority pseudo-reward) with genuine reward-design signal. Using a controlled tabular-GRPO simulator we derive an exact telescoping decomposition total = null + elicit + rd and measure each term across five prior-strength levels. The reward-design fraction of the naive estimator ranges from 0.139 at weak prior (ps=0.20) to 0.05 at strong prior (ps=0.80), with the elicitation term flipping sign at the self-consistency crossover. A pre-registered 2x2x2 factorial confirms non-additivity (interaction ratio 0.385; AxC effect -0.089). A points-vs-bounds pilot gate shows strong-prior regimes are point-identified while near-crossover regimes are only bounded. Re-audits of two named published results yield ELICITATION DOMINATED (elicitation share 0.98) and REWARD DESIGN DOMINATED (rd share 1.18) verdicts respectively, demonstrating the diagnostic value of the partition. We pre-commit to submit regardless of flip outcome; a non-flip is a finding of equal standing. We release a reusable one-command harness for any alignment paper to run the same audit.

AI agents rely on a harness of skills, tools, and workflows to solve complex problems. Continually improving this harness is essential for adapting to new tasks. However, existing optimization methods typically require ground-truth validation sets, yet such labeled data is difficult to acquire in practical deployment settings. To address this problem, we introduce Retrospective Harness Optimization (RHO), a self-supervised method that optimizes the agent harness using only past trajectories. Specifically, RHO selects a diverse coreset of challenging tasks from past trajectories and re-solves them in parallel. The agent analyzes these rollouts using self-validation and self-consistency, then generates candidate harness updates and selects the most effective one by its own pairwise self-preference. We evaluate RHO across three diverse domains, spanning software engineering, technical work, and knowledge work. Notably, a single optimization round improves the pass rate on SWE-Bench Pro from 59% to 78% without any external grading. Furthermore, our analysis demonstrates that RHO effectively targets prior failure modes. As a result, the optimized harness alters the agent's behavior patterns and sustains higher accuracy during long-horizon sessions.

Consistency training encourages models to behave similarly across different contexts, and has shown promise for reducing misalignment. We broaden the scope of consistency training in two ways. First, we introduce two new internal consistency targets: MLP Consistency Training (MLPCT), which matches post-activation MLP states, and Attention Consistency Training (AttCT), which matches per-head attention distributions. Second, we apply consistency training to four additional safety threats: persona in-context learning attacks, adversarial frustration, prefill attacks, and conditional misalignment. Across several models and threat settings, we find that consistency training reduces misalignment well beyond the sycophancy and jailbreak settings studied in prior work. We also find cases of cross-threat generalization, where training against one failure mode improves robustness to another, and identify a shared residual-stream mechanism underlying ACT, MLPCT, and AttCT, while distinguishing BCT as mechanistically distinct. Our results suggest that consistency training is a flexible and extensible framework for alignment, capable of unifying defenses against a broader class of model pathologies.

Self-consistency improves large language models by sampling multiple reasoning paths and selecting the most frequent answer, but majority voting often fails to recover correct answers that are already present among the samples. We address this limitation with Ranking-Improved Self-Consistency (RISC), which reformulates answer selection in self-consistency as a ranking problem. Instead of relying on a single uncertainty or confidence signal, RISC uses a lightweight LambdaRank model to score candidate answers with five carefully designed features that capture answer frequency, semantic centrality, and reasoning-trace consistency. We evaluate RISC on three datasets under a range of test-time budgets. Across datasets, RISC consistently achieves a better accuracy-efficiency trade-off than standard self-consistency and strong baselines, with particularly large gains on question answering benchmarks. Further analysis shows that the proposed features are individually useful and, more importantly, complementary, highlighting the value of learning to combine multiple informative signals for test-time answer selection.

In this report, we present our solution for Track 2 of the CVPR 2026 VidLLMs Challenge. This track evaluates visual relational reasoning in videos, where models must infer relations that are not always explicitly visible. We propose Answer Self-Consistency with Margin-Triggered Question Re-Arbitration (ASC-MQRA), a training-free test-time reasoning framework built on a multimodal reasoning model. The core ASC component performs multiple stochastic video question-answering runs and aggregates their answer choices through answer-level self-consistency. This substantially improves over single-pass inference and forms our final test submission. We further study MQRA, a conditional re-arbitration module for low-margin examples where the first-stage vote distribution indicates uncertainty. Our vote-margin analysis shows that low-margin examples often retain the ground-truth answer among the top candidates, motivating MQRA to narrow the candidate set and re-watch the video only over the retained candidates. On validation, MQRA further improves over ASC, indicating that low-margin vote distributions can provide a useful uncertainty signal. On test, however, MQRA slightly degrades performance relative to ASC, suggesting that re-arbitration is sensitive to the size and category distribution of the triggered subset. Our final test submission therefore uses ASC without re-arbitration, achieving 72.73 average accuracy and 78.34 category-wise macro average accuracy on validation, and 81.16 average accuracy and 80.91 category-wise macro average accuracy on test. This report details our prompting strategy, implementation setup, ablation studies, and diagnostic analyses. The code is available at https://github.com/data-analytics-labo/ASC-MQRA

Retrieval effectiveness varies substantially across queries, making it important to estimate ranking quality before relevance judgments are available. Query performance prediction (QPP) addresses this need, but most existing methods rely on external predictors after retrieval or reranking. In this paper, we study \textit{reranker-internal QPP}: can an LLM reranker estimate the quality of the ranking it has just produced? We investigate both training-free and training-based approaches. For training-free estimation, we examine metric-specific self-consistency across sampled rankings and verbalized confidence produced directly by the reranker. Experiments on TREC Deep Learning 2019--2022 with four LLMs show that self-consistency is competitive with the state-of-the-art (SOTA) approach and better calibrated in almost all settings, while direct verbalized confidence is severely overconfident. To improve verbalized confidence, we propose two supervised methods, Verb-Num and Verb-List, which enable LLM rerankers to produce calibrated ranking-quality estimates with only a few additional output tokens.

We describe our submission to the VRR Challenge @ CVPR 2026, built on the \emph{ImplicitQA} / \emph{VRR-QA} benchmark~\cite{implicitqa}: multiple-choice video question answering in which answers are deliberately \emph{not} observable in any single frame and must be inferred from spatial layout, motion, depth, viewpoint, causality, and social context across discontinuous frames of creative video. We conduct a systematic, training-free study spanning open-source Video-LMMs (Qwen2.5-VL~\cite{qwen25vl}, Qwen3-VL~\cite{qwen3vl}, InternVL3, Gemma-3, and the RL-tuned video reasoners Video-R1~\cite{videor1} and VideoChat-R1.5~\cite{videochatr15}) and a battery of inference-time strategies (chain-of-thought, question decomposition, describe-then-reason cascades, audio transcripts, spatial state prompting, self-consistency~\cite{selfconsistency}, multi-model ensembling, and category routing). Our central finding is that this benchmark is \emph{perception-bound rather than reasoning-bound}: reasoning-side augmentations are neutral-to-harmful, whereas base-model perceptual capability and lightweight test-time denoising are the only reliable levers. A per-category error analysis localizes the difficulty to low-level perception -- relative depth, viewpoint, and counting are the hardest categories, while causal and social reasoning are nearly solved -- and a prompt that explicitly injects monocular depth cues to attack the weakest category \emph{lowers} test accuracy by $5.8$ points, confirming that the model needs a better \emph{percept}, not a better \emph{procedure}.

Test-time reasoning has become a significant field of study since the introduction of chain-of-thought reasoning in large language models (LLMs). However, the mechanisms of this reasoning process are still under-explored -- from the same input prompt, and even the same partial solution, LLMs can produce varied answers if sampled multiple times. We propose to leverage question-asking as an inference-time intervention that articulates information about the model's hidden state. To achieve that, we present a student-teacher setting where a student asks questions to a teacher. We train a probe on the student's hidden state before and after asking a question and find it is predictive of the trajectory's final correctness, even before generating the teacher's answer. This suggests there is a meaningful signal from the self-diagnosis that occurs during question generation rather than information transfer from the teacher. We then frame question-asking as a sequential decision problem, using this probe as a quality score, and define a gating policy to ask questions that maximize likelihood of correctness. We find that the success of question-asking as an intervention is largely dependent on the model's self-consistency. Our empirical results show a gap between detection and recovery; while our gating policy captures model correctness and uncertainty, interventions are equally likely to harm correct trajectories as they are to recover incorrect ones. This gap between diagnosis and correction has broader implications on language models' capacity for self-refinement under uncertainty.

Inference-time sampling can elicit strong reasoning abilities from language models without additional training. Existing power-sampling methods do so by sharpening the distribution over full generated outputs, favoring completions that are individually likely under the model. We argue that this is the wrong object to target for reasoning: a completion entangles a reasoning trace with a final answer, whereas what matters is whether an answer is supported by many plausible reasoning paths. We therefore shift the target from the full-output distribution to the sharpened answer marginal, making self-consistency an inference-time objective rather than a post-hoc voting criterion. Surprisingly, this marginal target admits an efficient approximation: we propose a simple, purely autoregressive parallel sampling algorithm that approximately samples from the sharpened answer marginal, eliciting stronger performance than standard power sampling on mathematics and coding benchmarks while being orders of magnitude faster.

Reinforcement learning with evaluation metrics as rewards is widely used to enhance specific capabilities of language models. However, for tasks such as factually consistent summarisation, existing metrics remain underdeveloped, limiting their effectiveness as signals for shaping model behaviour.While individual factuality metrics are unreliable, their combination can more effectively capture diverse factual errors. We leverage this insight to introduce an automated training pipeline that improves factual consistency in summaries by aggregating scores from different weak metrics. Our approach avoids the need for complex reward shaping by mapping scores to preferences and filtering out cases with high disagreement between metrics. For each source document, we generate lexically similar summary pairs by varying decoding strategies, enabling the model to learn from factual differences caused by subtle lexical differences. This approach constructs a high-quality preference dataset using only source documents.Experiments demonstrate consistent factuality gains across models, ranging from early encoder-decoder architectures to modern large language models, with smaller models reaching comparable factuality to larger ones.

Chain-of-thought (CoT) reasoning with self-consistency improves performance by aggregating multiple sampled reasoning paths. In this setting, correctness is no longer tied to a single reasoning trace but to the aggregation rule over a pool of candidate paths, making aggregation uncertainty the central challenge. This issue is critical where confidently incorrect answers are far more costly than abstentions. We introduce a conformal procedure for CoT reasoning that directly addresses aggregation uncertainty. Our approach replaces majority voting with weighted score aggregation over reasoning paths and calibrates an abstention rule using conformal risk control. This approach leads to finite-sample guarantees on the confident-error rate--the probability that the system answers and is wrong. We further identify score separability as the key condition under which abstention provably improves selective accuracy, and derive closed-form expressions that predict accuracy gains from calibration data alone. The method is fully inference-time, and requires no retraining. Across four benchmarks, four open-source models, and three score classes, realized confident-error rates are consistent with the prescribed targets up to calibration-split and test-set variability. Our method achieves $90.1\%$ selective accuracy on GSM8K by abstaining on less than $5\%$ of problems, compared with $82\%$ accuracy under majority-voting baseline.

BACKGROUND: Coding Motivational Interviewing (MI) sessions is essential for understanding client behaviors and predicting outcomes, but it requires substantial time and labor from trained MI professionals. Recent advances in audio-language models (ALMs) offer new opportunities to automate MI coding by capturing multimodal behavioral signals. OBJECTIVE: This study aims to develop an automatic MI coding approach based on ALMs that analyzes raw audio input and integrates predictions from multiple reasoning trajectories using self-consistency to improve coding robustness. METHODS: We experimented with five recorded sessions from de-identified MI audio tapes. We deployed ALMs with four complementary analytic prompts to support utterance-level reasoning: analytic prompting for verbal cues, prosody-aware prompting for acoustic cues, evidence-scoring prompting for quantitative hypothesis testing, and comparative prompting for contrastive reasoning. Three stochastic samples were drawn for each prompt, generating 12 independent reasoning trajectories per utterance. Final predictions were determined by majority voting across all trajectories. RESULTS: Performance was evaluated using accuracy, precision, recall, and macro-F1 scores. The proposed multimodal self-consistency approach achieved 52.56% accuracy, 54.03% precision, 47.45% recall, and a macro-F1 score of 46.40%, exceeding baseline methods. Systematic ablation experiments that removed individual modules consistently degraded performance on the primary metrics. CONCLUSIONS: Multimodal self-consistency outperforms single-pass baseline prompting approaches for MI coding. These findings suggest that incorporating both what clients say and how they say it can support more reliable automatic MI coding.

State-of-the-art physical AI models generate a chunk of actions per inference through diffusion or flow matching, iteratively refining an initial noise sample into an action trajectory. Because this inference process is inherently stochastic, committing to a single trajectory per round is brittle, and this brittleness compounds across the many sequential rounds that comprise a complete episode. We introduce KeyStone, an inference-time self-consistency method for diffusion-based action generation that draws $K$ candidate action chunks in parallel from a shared model context, clusters them in continuous action space, and returns the medoid of the largest cluster -- no additional model required. Two properties make this practical. First, the compact nature of action trajectories makes diffusion inference memory-bandwidth bound, leaving spare compute capacity to run $K$ chains in parallel with no additional wall-clock latency. Second, unlike token or pixel spaces where distance carries no semantic meaning and selection requires a learned judge, action chunks are geometrically structured such that Euclidean distance directly reflects physical similarity, making selection principled and judge-free. Across diverse vision-language-action models (VLAs) and world-action models (WAMs), KeyStone improves task success rates by up to \textbf{13.3%} over single-trajectory sampling with negligible latency overhead, while having on par accuracy with model-based selectors at no training cost. We open source KeyStone at https://github.com/dywsjtu/keystone.

Self-play red team is an established approach to improving AI safety in which different instances of the same model play attacker and defender roles in a zero-sum game, i.e., where the attacker tries to jailbreak the defender; if self-play converges to a Nash equilibrium, the model is guaranteed to respond safely within the settings of the game. Although the parameter sharing enforced by the use of the same model for the two roles improves stability and performance, it introduces fundamental theoretical and architectural limitations. We show that the set of Nash equilibria that can be reached corresponds to a broad class of behaviours that includes trivial always refuse strategies and oracle-like defenders, thus limiting practical applicability. We then show that when attacker and defender share and update the same base model, the dynamics collapse to self-consistency, so that attacks do not enforce adversarial pressure on the defender. In response, we propose Anchored Bipolicy Self-Play, which trains distinct role-specific LoRA adapters on top of a frozen base model, thereby maintaining stable optimisation while preserving adversarial pressure through explicit role separation. In relation to standard self-play, we show up to 100x greater parameter efficiency than finetuning and consistent improvements in safety compared to self-play fine-tuned models. We evaluate on Qwen2.5-{3B, 7B,14B}-IT models across widely used safety benchmarks, showing improved robustness without loss of reasoning ability. Cross-play experiments further show that our attacker and defender models are superior to self-play in terms of adversarial defence and safety.

A standard technique for scaling inference-time reasoning is Self-Consistency, whereby multiple candidate answers are sampled from an LLM and the most common answer is selected. More recently, it has been shown that weighted majority voting (e.g. Confidence-Informed Self Consistency (CISC)), which assigns a confidence value to each candidate answer and chooses the answer with the largest accumulated score, tends to be more accurate on a wide range of popular benchmarks. In practice, weighted majority voting necessitates calling a critic LLM on each candidate's reasoning trace to produce the answer's confidence score. This secondary series of LLM calls greatly increases the overhead and cost of weighted majority voting, despite its potential performance benefits. To reduce this expense, we propose VecCISC, a lightweight, adaptive framework that uses a measure of semantic similarity to filter reasoning traces that are semantically equivalent to others, degenerate, or hallucinated, thus decreasing the number of candidate answers that must be evaluated by the critic. To ensure adequate experimental thoroughness, we evaluate VecCISC on five challenging, widely-adopted datasets spanning the domains of mathematics, chemistry, biology, commonsense reasoning, and the humanities. Our results demonstrate that VecCISC reduces the total token usage by 47%, while maintaining or exceeding the accuracy of CISC.

Large Language Models often improve accuracy on reasoning tasks by sampling multiple Chain-of-Thought (CoT) traces and aggregating them with majority voting (MV), a test-time technique called self-consistency. When we truncate a CoT partway through and regenerate the remainder, we observe that traces with correct answers reproduce their original answer more often than traces with wrong answers. We use this difference as a reliability signal, prefix consistency, that weights each candidate answer by how often it reappears under regeneration. It requires no access to token log-probabilities or self-rating prompts. Across five reasoning models and four math and science benchmarks, prefix consistency is the best correctness predictor in most settings, and reweighting votes by it reaches Standard MV plateau accuracy at up to 21x fewer tokens (median 4.6x). Our code is available at https://github.com/naoto-iwase/prefix-consistency.

Self-consistency detects hallucinations by generating multiple sampled answers to a question and measuring agreement, but this requires repeated decoding and can be sensitive to lexical variation. Semantic self-consistency improves this by clustering sampled answers by meaning using natural language inference, but it adds both sampling cost and external inference overhead. We show that first-token confidence, phi_first, computed from the normalized entropy of the top-K logits at the first content-bearing answer token of a single greedy decode, matches or modestly exceeds semantic self-consistency on closed-book short-answer factual question answering. Across three 7-8B instruction-tuned models and two benchmarks, phi_first achieves a mean AUROC of 0.820, compared with 0.793 for semantic agreement and 0.791 for standard surface-form self-consistency. A subsumption test shows that phi_first is moderately to strongly correlated with semantic agreement, and combining the two signals yields only a small AUROC improvement over phi_first alone. These results suggest that much of the uncertainty information captured by multi-sample agreement is already available in the model's initial token distribution. We argue that phi_first should be reported as a default low-cost baseline before invoking sampling-based uncertainty estimation.

Advances in inference methods have enabled language models to improve their predictions without additional training. These methods often prioritize raw performance over cost-effective compute usage. However, computational efficiency is key for real-world applications with resource constraints. We provide a systematic analysis of the inference scaling strategies self-consistency, self-refinement, multi-agent debate, and mixture-of-agents, to study their computational performance tradeoffs. We evaluate methods on two reasoning benchmarks (MMLU-Pro, BBH) and include extensive parameter configurations (e.g., scaling the number of parallel predictions, agents, and debate rounds) across different model sizes. Across 34 configurations and over 100 evaluations, we compute the Pareto-optimal front to select methods that achieve the best accuracy with the lowest computational budget. Notably, inference scaling improves accuracy by up to +7.1% points over chain-of-thought at the highest evaluated budgets (20x the CoT compute budget) on MMLU-Pro. With an equal computing budget, debate and mixture-of-agents outperform self-consistency by 1.3% and 2.7% points, respectively. While self-consistency saturates earlier, multi-agent gains persist, particularly on more complicated tasks. We identify a simple multi-agent design guideline: mixture-of-agents is most efficient when the number of parallel generations exceeds the number of sequential aggregations.

Small instruct-tuned LLMs produce degenerate verbal confidence under minimal elicitation: ceiling rates above 95%, near-chance Type-2 AUROC, and Invalid validity profiles. We test whether confidence-conditioned supervised fine-tuning (CSFT) with self-consistency-derived targets can close the gap between internal information and verbal readout. A pre-registered Phase 0 protocol on Gemma 3 4B-it with a modal filter restricting training to items with correct modal answers produced a negative result: AUROC2 dropped from 0.554 to 0.509 due to label-entropy collapse in the training targets. An exploratory rescue removed the filter, training on all 2,000 calibration items. This produced a binary verbal correctness discriminator with AUROC2 = 0.774 on held-out TriviaQA, compressing a 10-sample self-consistency signal (AUROC2 = 0.999) into a single-pass readout exceeding logit entropy (0.701). The shuffled-target control showed no improvement (0.501). On MMLU, accuracy improved from 54.2% to 77.4% with the shuffled model at baseline (56.1%), supporting a target-dependent interpretation. The result is exploratory, binary rather than continuously calibrated, and observed at a single scale. It identifies two design lessons: confidence training requires label entropy, and correct targets regularise output format.

Graphical User Interface (GUI) grounding requires mapping natural language instructions to precise pixel coordinates. However, due to visually homogeneous elements and dense layouts, models typically grasp semantic intent yet struggle with achieving precise localization. While scaling sampling attempts (Pass@k) reveals potential gains, static self-consistency strategies derived from geometric clustering often yield limited improvements, as the model's predictions tend to be spatially dispersed. In this paper, we propose replacing static consistency strategies with a learnable selection mechanism that selects the optimal target by critiquing its own proposals rendered on the screenshot. Given the significant disparity between the model's grounding and critiquing capabilities, we propose a co-evolving Propose-then-Critic framework. To jointly optimize these, we introduce a maturity-aware adaptive co-evolutionary reinforcement learning paradigm. This approach dynamically balances the training objectives of proposer and critic, where the diversity of the proposer's outputs enhances critic robustness, while the critic's maturing discrimination capability conversely unlocks the proposer's potential for extensive spatial exploration, fostering the mutual reinforcement and co-evolution of both capabilities, thereby ensuring generalizability to adapt to diverse and complex interface layouts. Extensive experiments over 6 benchmarks show that our method significantly enhances both grounding accuracy and critic reliability.

While self-consistency is known to improve performance on symbolic reasoning, its effect on the recall of encyclopedic knowledge is unclear due to a lack of targeted evaluation grounds. To address this, we establish such a knowledge recall split for the popular MMLU benchmark by applying a data-driven heuristic from prior work. We validate this split by showing that the performance patterns on the symbolic reasoning and knowledge recall subsets mirror those of GSM8K and MedMCQA, respectively. Using this solid ground, we find that self-consistency consistently improves performance across both symbolic reasoning and knowledge recall, even though its underlying CoT prompting is primarily effective for symbolic reasoning. As a result, we achieve an 89% accuracy on MMLU, the best performance to date with the use of GPT-4o.

Self-consistency (SC) is a popular technique for improving the reasoning accuracy of large language models by aggregating multiple sampled outputs, but it comes at a high computational cost due to extensive sampling. We introduce a hybrid ensembling approach that leverages the complementary strengths of two distinct modes of reasoning: Chain-of-Thought (CoT) and Program-of-Thought (PoT). We describe a general framework for combining these two forms of reasoning in self-consistency, as well as particular strategies for both full sampling and early-stopping. We show that CoT-PoT ensembling not only improves overall accuracy, but also drastically reduces the number of samples required for SC by a factor of 9.3x. In particular, the majority of tasks (78.6%) can be addressed with only two samples, which has not been possible with any prior SC methods.

Large language models (LLMs) often produce confident yet incorrect responses, and uncertainty quantification is one potential solution to more robust usage. Recent works routinely rely on self-consistency to estimate aleatoric uncertainty (AU), yet this proxy collapses when models are overconfident and produce the same incorrect answer across samples. We analyze this regime and show that cross-model semantic disagreement is higher on incorrect answers precisely when AU is low. Motivated by this, we introduce an epistemic uncertainty (EU) term that operates in the black-box access setting: EU uses only generated text from a small, scale-matched ensemble and is computed as the gap between inter-model and intra-model sequence-semantic similarity. We then define total uncertainty (TU) as the sum of AU and EU. In a comprehensive study across five 7-9B instruction-tuned models and ten long-form tasks, TU improves ranking calibration and selective abstention relative to AU, and EU reliably flags confident failures where AU is low. We further characterize when EU is most useful via agreement and complementarity diagnostics.

Open-weight Small Language Models(SLMs) can provide faster local inference at lower financial cost, but may not achieve the same performance level as commercial Large Language Models (LLMs) that are orders of magnitudes larger. Consequently, many of the latest applications of LLMs, such as software engineering agents, tend to be evaluated on larger models only, leaving the issue of improving the cost-benefit trade-off of such applications neglected. This paper proposes Atropos, a predictive early-termination analysis and hotswap technique that aims to improve the cost-benefit trade-off for LLM-based agents that use self-consistency. The core component of ATROPOS is a predictive model based on structural properties of LLM inferences: after merging multiple agentic inference paths into a graph representation, ATROPOS uses Graph Convolutional Network (GCN) to predict whether an ongoing inference will eventually succeed or not. If an agentic task instance running on the source LLM is predicted to fail, ATROPOS subsequently performs hotswapping, i.e., migrating the on-going inference context onto the more capable target LLM: this is feasible because LLM contexts are stateless. An empirical evaluation of ATROPOS using three recent LLM-based agents shows that ATROPOS can predict early termination of eventually failing inferences with the accuracy of 0.85 at the midpoint of the inference. Hotswapping LLMs for such inferences can convert up to 27.57% of them to be successful. Consequently, ATROPOS achieves 74.35% of the performance of closed LLMs with as low as only 23.9% of the cost.

Training process reward models (PRMs) requires step-level correctness labels, obtained either through expensive human annotation or by relying on ground-truth answers, limiting the ability to scale process-level supervision. We propose ScalePRM, which scales verification compute as an alternative: given a problem and a candidate solution, we generate multiple independent verifications of each reasoning step and aggregate their judgments to produce synthetic step-level labels without ground truth. We explore two representative inference-time scaling strategies, parallel scaling through self-consistency and sequential scaling through meta-critique, and train generative PRMs on the resulting synthetic data. On ProcessBench, a benchmark for identifying erroneous steps in mathematical reasoning, PRMs trained on step-level self-consistency data achieve 67.5 F1, surpassing reference-guided training with ground-truth access (66.4 F1) and GPT-4o as a critic (61.9 F1). When deployed as reward signals in RL training with Qwen2.5-Math-7B, our best PRM achieves 47.4% average accuracy across six mathematical reasoning benchmarks, outperforming ground-truth-based RLVR (43.9%). We also identify and address reward exploitation patterns unique to generative PRM-based RL. Our results demonstrate that scaling verification compute is a viable alternative to ground-truth supervision for training process reward models.

Self-consistency (SC) is a widely used test-time inference technique for improving performance in chain-of-thought reasoning. It consists of generating multiple responses, or ``samples", from a large language model (LLM) and selecting the most frequent answer. This procedure can naturally be viewed as a majority vote or empirical mode estimation. Despite its effectiveness, self-consistency is prohibitively expensive at scale when naively applied to datasets, and it lacks a unified theoretical understanding of sample efficiency and scaling behavior. In this paper, we provide the first comprehensive analysis of SC's scaling behavior and its variants, drawing on mode estimation and voting theory. We derive and empirically validate power law scaling for self-consistency across datasets, and analyze the sample efficiency for fixed-allocation and dynamic-allocation sampling schemes. From these insights, we introduce Blend-ASC, a novel variant of self-consistency that dynamically allocates samples to questions during inference, achieving state-of-the-art sample efficiency. Our approach uses 4.8 times fewer samples than vanilla SC on average, outperforming both fixed- and dynamic-allocation SC baselines, thereby demonstrating the superiority of our approach in terms of efficiency. In contrast to existing variants, we note that Blend-ASC is hyperparameter-free, supports batching, and can fit any budget of samples, ensuring it can be easily applied to any self-consistency application.

Large Language Models (LLMs) often hallucinate, limiting their reliability in sensitive applications. In black-box settings, several self-consistency-based techniques have been proposed for hallucination detection. We empirically show that these methods perform nearly as well as a supervised (black-box) oracle, leaving limited room for further gains within this paradigm. To address this limitation, we explore cross-model consistency checking between the target model and an additional verifier LLM. With this extra information, we observe improved oracle performance compared to purely self-consistency-based methods. We then propose a budget-friendly, two-stage detection algorithm that calls the verifier model only for a subset of cases. It dynamically switches between self-consistency and cross-consistency based on an uncertainty interval of the self-consistency classifier. We provide a geometric interpretation of consistency-based hallucination detection methods through the lens of kernel mean embeddings, offering deeper theoretical insights. Extensive experiments on QA-style hallucination detection benchmarks show that this approach maintains high detection performance while significantly reducing computational cost.