NeurIPS 2026 Papers

When language models reason in chain-of-thought or collaborate in pipelines, they serialize structured information into natural language. How much compositional structure survives this bottleneck? We propose a round-trip protocol that answers this question empirically for tree-structured symbolic expressions. A generator converts a procedurally generated arithmetic expression into a word problem, a separate extractor recovers the expression from the word problem alone, and symbolic equivalence provides an exact oracle. Evaluating all pairwise combinations of fifteen models yields a communication matrix whose marginals separate generation quality from extraction quality. Three main findings emerge. First, the channel is lossy and asymmetric: swapping which model generates and which extracts shifts accuracy by up to 53.2 points, and the best pair reaches 82.0\% by combining different models on each end rather than the same model on both. Second, at least 84\% of round-trip failures originate at generation, and difficulty is driven by tree structure (operator count, depth, right-branching) rather than model family. Third, the channel is trainable: $\sim3600$ fine-tuning examples that share the evaluation's operators and tree shapes lift every open-weight model above untrained GPT-5, an upper bound under matched semantics. A disjoint-domain regime with new operators and vocabulary also raises every open-weight model, confirming the gain is not an artifact of matched semantics, though a gap to the frontier remains. Together these results identify compositional serialization as a primary limiting factor when models communicate hierarchical structure through natural language.

Group Relative Policy Optimization (GRPO) is widely used to post-train reasoning language models, and its rollout group size $G$ is typically treated as a compute hyperparameter. We argue that $G$ also sets a difficulty threshold below which prompts produce no learning signal. Under binary verifier rewards, a rollout group only produces a learning signal when it contains both a success and a failure, and the rate at which a prompt yields such mixed groups turns on sharply near $p_c\simeq\ln 2/G$. Below that threshold the prompt is invisible to the relative-advantage estimator regardless of optimizer state. We verify this gate directly in pass@256 rollout traces from Qwen3-1.7B, Qwen3-8B, and Llama-3.1-8B, where empirical mixed-group rates match the binomial prediction within $1.4$–$1.9$ percentage points on average per model. Sweeping fixed-$G$ GRPO with $G\in\{8,16,32,64,128\}$ then shows that difficulty-binned per-prompt improvement after training tracks the frontier shape with binned $R^2$ between $0.962$ and $0.984$ across all three model families. Treating $p_c$ as a moving target, we instantiate a frontier-tracking group-size schedule that improves the best fixed-$G$ baseline by $4.2$–$5.3$ points on MATH-500 and $2.4$–$3.4$ points on GSM8K under matched rollout budget. Ablations attribute the gain primarily to frontier-aware switch timing.

Test-time scaling (TTS) improves language model outputs by spending additional inference compute — generating multiple candidates, searching over partial sequences, or iteratively refining drafts. These techniques yield large gains on mathematics and code, but have been developed and stress-tested almost exclusively on tasks where verification is straightforward. We conduct the first compute-normalised comparison of five TTS families across five open-ended generation benchmarks spanning medicine, law, finance, general chat, and creative writing — grounded in a unified framework that decomposes the effectiveness of each method's token budget into exploration and exploitation. The answer depends on which side of that decomposition you examine. Scaling exploration works: the best candidate in the pool improves steadily with compute across all settings. What breaks is exploitation — the step that converts a rich candidate pool into a final output. With state-of-the-art generators, reward models correlate at only $\hat{\rho}_v$ $\approx$ 0.12 with true quality, rendering selection near-random regardless of budget. Tree search amplifies this failure through diversity collapse. Refinement helps on one of five benchmarks; its apparent gains elsewhere are confounded. Only synthesis across candidates (Fusion) consistently improves over single-sample baselines, yet still recovers only ${\sim}$40% of available quality.The candidate pool is not the bottleneck — choosing from it is.

Test-Time Prompt Tuning (TPT) adapts large pretrained models to unseen target distribution shifts by updating only lightweight prompt tokens. Existing TPT methods are effective, but they mainly optimize each prompt update locally. Across test-time steps, prompts updated at the current step are reused for later predictions and further updates, so noisy or biased unlabeled objectives can accumulate and destabilize the prompt-update trajectory. We identify two failure modes: magnitude expansion, where updates move the prompt too far, and directional drift, where consecutive updates point in inconsistent directions. To analyze and control these failures, we propose Test-Time Prompt-Agnostic Decomposition ($\mathtt{TPD}$), which characterizes observed prompt-update trajectories from a dynamical-systems view and decomposes base prompt updates into magnitude and direction. $\mathtt{TPD}$ computes a spectral radius to measure update expansion and decomposes update directions into persistent, oscillatory, and residual components. Building on this decomposition, Adaptive Koopman Control ($\mathtt{AKC}$) regulates update magnitude by shrinking updates under expansive recent dynamics, while Hankel Update Router ($\mathtt{HUR}$) refines update direction by preserving persistent components and suppressing oscillatory and residual components. Together, $\mathtt{AKC}$ and $\mathtt{HUR}$ produce a stabilized prompt update. As a prompt-agnostic framework, $\mathtt{TPD}$ can be plugged into visual and text TPT methods without modifying the backbone, prompt architecture, or adaptation loss. Experiments on 15 datasets show that $\mathtt{TPD}$ consistently improves 10 baselines, with accuracy gains of 3-8\% for visual prompts and over 2\% for text prompts.

Graph anomaly detection (GAD) is critical to many real-world systems. However, supervised GAD models often degrade under distribution shifts due to the closed-world training assumption. Test-time adaptation (TTA) offers a promising solution by updating pre-trained models with unlabeled test data. Yet, applying TTA to GAD poses a unique dilemma. High-confidence samples provide reliable pseudo-labels, but they are scarce and biased toward low-shift regions. Low-confidence samples better reflect target shifts, but they are too noisy for direct cross-entropy optimization and often require label-free objectives that are misaligned with anomaly discrimination. To address this dilemma, we propose SADOS, a Shifted Augmentation-based Dynamic Objective Scheduling framework for test-time GAD. SADOS turns unreliable target samples into drift-carrying signals for reliable supervised adaptation. It distills prototypes from confidence-stratified unreliable samples and injects them into reliable samples to generate label-preserving yet shift-oriented augmentations. This expands coverage over shifted target regions. In parallel, SADOS dynamically schedules the adaptation objective. It uses contrastive learning to capture early-stage drift signals, and gradually anneals its weight so that classification-oriented anomaly detection dominates later adaptation. Extensive experiments demonstrate that SADOS robustly improves test-time GAD under distribution shifts.

Test-time compute has emerged as a promising strategy to enhance the reasoning abilities of large language models (LLMs). However, this strategy has in turn increased how much users pay cloud-based providers offering LLM-as-a-service, since providers charge users for the amount of test-time compute they use to generate an output. In our work, we show that the market of LLM-as-a-service is socially inefficient–providers have a financial incentive to increase the amount of test-time compute, even if this increase contributes little to the quality of the outputs. To address this inefficiency, we introduce a reverse second-price auction mechanism where providers bid their offered price and (expected) quality for the opportunity to serve a user, and users pay proportionally to the marginal value generated by the winning provider relative to the second-highest bidder. To illustrate and complement our theoretical results, we conduct experiments with multiple instruct models from the $\texttt{Llama}$ and $\texttt{Qwen}$ families, as well as reasoning models distilled from $\texttt{DeepSeek-R1}$, on benchmark datasets covering mathematics, science, and open-ended question answering.

Test-time training (TTT) enhances model performance by explicitly updating designated parameters prior to each prediction to adapt to the test data. While TTT has demonstrated considerable empirical success, its theoretical underpinnings remain limited, particularly for nonlinear models. In this paper, we investigate the combination of TTT with in-context learning (ICL), where the model is given a few examples from the target distribution at inference time. We analyze this framework in the setting of single-index models, where the feature vector is drawn from a hidden low-dimensional subspace. For single-layer transformers trained with gradient-based algorithms and adopting TTT, we establish an upper bound on the prediction risk. Our theory reveals that TTT enables the single-layer transformers to adapt to both the feature vector and the link function, which vary across tasks. This creates a sharp contrast with ICL alone, which is theoretically difficult to adapt to shifts in the link function. Moreover, we provide the convergence rate with respect to the data length, showing the predictive error can be driven arbitrarily close to the noise level as the context size and the network width grow.

Language models are increasingly deployed alongside test-time search. We argue that we should therefore shift the role of RL post-training from converging on a single best response to producing a diverse pool of competent candidates. Standard policy gradient methods such as GRPO optimize a fixed scalar reward and drive the policy toward near-duplicate responses, erasing the diversity search needs. We propose Vector Policy Optimization (VPO), which exploits the fact that practical rewards are vector-valued, e.g., per-test-case correctness, per-criterion ratings, or per-hop credit. Instead of collapsing these into one scalarization, VPO samples randomized scalarizations from a distribution over the reward simplex, incentivizing candidates to specialize to different trade-offs along the Pareto front. The model emits multiple candidates per prompt, with each conditioning on the previous ones, allowing for directed diversification. VPO is a drop-in replacement for the GRPO advantage estimator. Across four domains, VPO consistently improves test-time best@$k$ over scalar baselines, with gains widening as the test-time budget grows.

While test-time scaling has revolutionized reasoning in large language models, generative video reasoning remains bottlenecked by a single-shot paradigm. We demonstrate that searching over denoising steps cannot rescue logically flawed rollouts because spatial trajectories commit early in the diffusion process. Root-level Best-of-$N$ (BoN) sampling is similarly inefficient: reasoning errors cluster early in the temporal axis, and resampling blindly discards verified upstream progress. To unlock effective test-time scaling for video models, we introduce \textbf{Temporal Backtracking Search (TBS)}, which shifts the search space to the temporal axis. TBS transforms video generation into an iterative generate--verify--restart loop via three core mechanisms: (1) \textit{variable-$K$ conditioning} to resume generation from arbitrary clean prefixes; (2) \textit{temporal process verification} to localize failures and extract valid restart anchors; and (3) \textit{prefix-based search} to reallocate compute toward extending correct trajectories rather than root resampling. Across algorithmic, navigation, and robotics domains, TBS Pareto-dominates matched-budget BoN. In a strict out-of-distribution setting where one-shot generation collapses ($0.7\%$ for BoN), TBS achieves $22.7\%$, with every solved episode stemming from a restarted branch. Ultimately, TBS reveals that the local reasoning competence of video models far exceeds what single-shot rollouts indicate, providing a scalable test-time framework to unlock it.

End-to-end autonomous driving systems are known to suffer from shortcut learning and causal confusion, limiting robust generalization under distribution shift. To address this, recent Vision-Language Model (VLM) approaches introduce intermediate linguistic reasoning steps including chain-of-thought, chain-of-causation, and counterfactual (CF) reasoning, and report consistent improvements in closed-loop driving. Among these, CF reasoning has been emphasized as enabling self-reflective, high-level causal understanding, yet whether these gains reflect genuine causal reasoning or language-induced distributional bias remains unclear. In this work, we examine what CF reasoning actually induces in VLM-based autonomous driving using accident scenarios with explicit causal structure and a discrete meta-action space. We observe that CF reasoning consistently shifts the action distribution toward more conservative behaviors, accompanied by increased perplexity and policy entropy, and that this effect persists under reinforcement fine-tuning, suggesting a structural rather than transient artifact. However, visual attention analysis and causal intervention experiments show that CF reasoning does not meaningfully alter visual grounding, and VLMs respond similarly to causal and non-causal perturbations, unlike humans who selectively react to causal factors. Overall, these results suggest that CF reasoning does not induce genuine causal reasoning grounded in visual structure. Instead, it acts as a linguistic mechanism that reshapes the action distribution, teaching models what to say but not how to reason, highlighting a fundamental gap between language-based reasoning and causal understanding in VLMs.

Large language models (LLMs) often fail to reason under temporal cutoffs: when prompted to answer from the standpoint of an earlier time, they exploit knowledge that became available only later. We study this failure through the lens of ex-ante reasoning, where a model must rely exclusively on information knowable before a cutoff. Through a systematic analysis of prompt-level interventions, we find that temporal leakage is highly sensitive to cutoff formulation and instruction placement: explicit cutoff statements outperform implicit historical framings, and prefix constraints reduce leakage more effectively than suffix constraints. These findings indicate that prompting can steer models into a temporal frame, but does not endow them with the ability to verify whether a response is temporally admissible. We further argue that supervised fine-tuning is insufficient, since ex-ante correctness is not an intrinsic property of an answer, but a relation between the answer and the cutoff. To address this gap, we propose \textbf{TCFT}, a \underline{T}emporal \underline{C}ritique \underline{F}ine-\underline{T}uning framework that trains models to acquire cutoff-aware temporal verification. Given a query, a cutoff, and a candidate response, TCFT teaches the model to identify post-cutoff leakage, explain temporal boundary violations, and judge temporal admissibility. Experiments with Qwen2.5-7B-Instruct and Qwen2.5-14B-Instruct show that TCFT consistently outperforms prompting and SFT baselines, reducing average leakage by 41.89 and 37.79 percentage points, respectively.

Chest X-ray visual question answering (CXR VQA) requires models not only to predict correct answers, but also to produce reliable medical reasoning. However, existing reinforcement-learning-based training typically relies on answer-level rewards, which are often too coarse to improve chain-of-thought (CoT) quality and can become ineffective when group-level advantages collapse to zero. We propose \textbf{Teach-to-Reason (T2R)}, a framework that introduces comparison-based supervision into CoT optimization through a self-improving \emph{Teacher} and a competition-guided \emph{Reasoner}. As the Teacher is iteratively strengthened via self-competition, the Reasoner is optimized against progressively stronger Teacher-generated references. We further introduce a case-wise reward design that preserves the original reward-induced positive/negative partition when it is informative, and restores supervision from competition scores when the original reward signal degenerates. Experiments on multiple CXR open-ended VQA benchmarks show that T2R consistently outperforms strong baselines, indicating that comparison-based supervision, when integrated in a controlled and principled manner, provides a more effective training signal for reasoning optimization.

Alignment faking (AF) occurs when an LLM strategically complies with training objectives to avoid value modification, reverting to prior preferences once monitoring is lifted. Current detection methods focus on conversational settings and rely primarily on Chain-of-Thought (CoT) analysis, which provides a reliable signal when strategic reasoning surfaces, but cannot distinguish deception from capability failures if traces are absent or unfaithful. We formalize AF as a composite behavioural event and detect it through observable tool selection, where the LLM selects the safe tool when unmonitored, but switches to the unsafe tool under monitoring that rewards helpfulness over safety, while its reasoning still acknowledges the safe choice. We release a dataset of 108 enterprise IT scenarios spanning Security, Privacy, and Integrity domains under Corruption and Sabotage pressures. Evaluating six frontier LLMs across five independent runs, we find mean AF detection rates between 3.5\% and 23.7\%, with vulnerability profiles varying by domain and pressure type. These results suggest that susceptibility reflects training methodology rather than capability alone.

In real home deployments, household agents must often operate from a complete household scene and a situated household request, rather than from a clean task specification. Such requests require agents to identify task-relevant entities, recover intended task conditions, and resolve ordering constraints from the surrounding scene context. We formalize this capability as full-scene household reasoning: given a complete household scene and a situated household request, an agent must infer executable task structure before producing a grounded skill-level action sequence. This setting is challenging because complete household scenes contain substantial task-irrelevant information, making direct complete-scene prompting inefficient and error-prone. In practical deployment, this challenge is further amplified by privacy and local compute constraints, which favor compact open-weight models with limited long-context reasoning ability. We propose TaskGround, a training-free and model-agnostic Ground-Infer-Execute framework that grounds complete scenes into compact task-relevant scene slices, infers executable task structure, and compiles it into grounded skill-level action sequences. To evaluate this setting, we introduce FullHome, a human-validated evaluation suite of 400 household tasks spanning diverse home-scale environments and both goal-oriented and process-constrained requirements. On FullHome, TaskGround improves task success rates by large margins across both proprietary and open-weight models. Notably, it makes Qwen3.5-9B competitive with GPT-5 under direct complete-scene prompting while reducing total input-token cost by up to 18x. Our results identify executable task-structure inference as a central bottleneck in full-scene household reasoning and show that structured grounding can make compact local models substantially more effective for practical household deployment.

Reinforcement learning (RL) has substantially improved the reasoning capabilities of vision-language models (VLMs), yet it often triggers chain-of-thought (CoT) obfuscation, a regime where models achieve high accuracy while producing reasoning traces that are ungrounded and difficult to monitor. While prior work documents this decay at the behavioral level, the underlying mechanistic drivers of such representational drift remain poorly understood. In this work, we identify that obfuscation is a representational pathology, where RL-induced optimization causes task-agnostic template features to replace visually grounded content in the model’s activation space. Based on this insight, we propose TAME, an activation-level intervention framework that leverages Sparse Autoencoders (SAEs) to regularize VLM internal features during RL training. Specifically, TAME utilizes an LLM-as-judge monitor to detect behavioral obfuscation and maps these signals to specific internal directions, penalizing template intrusion while preserving visual evidence. By directly intervening on the features responsible for obfuscation, our method restores the transparency and visual grounding of reasoning traces. Experiments on the VIRL-39k and SPA-VL benchmarks across two model families show that TAME improves CoT monitorability (by up to +30.9% on VIRL and +16.7% on SPA-VL over GRPO), while preserving task accuracy and general-capability performance across four standard VLM benchmarks.

Decoder-only Large View Synthesis Models (LVSMs) with KV-cache have recently achieved state-of-the-art quality by regressing novel views from neural networks without reconstructing 3D geometry, while Test-Time Training (TTT) layers can be introduced to save computation, but sacrifice quality. We propose **T**est-**T**ime **B**aking (**TTB**), a strategy that learns to bake global scene information from cross-view attention into a lightweight MLP, enabling fast and high-fidelity novel-view decoding. The first key is to repurpose the fast-weight MLP of TTT to learn the cross-view attention mapping explicitly from its input-output pairs, rather than replacing attention entirely, achieving the same rendering efficiency as TTT while delivering substantially higher quality. The second key addresses a modality misalignment between context prefilling and novel-view rendering: we disentangle context tokens into camera-only context rendering tokens to serve as TTB input, and context ground-truth tokens with image content to serve as the TTB mapping target, ensuring the baked MLPs are relieved of the modality generalization challenge between context and novel-view processing. Our method achieves up to **23x** rendering speedup with quality on par with the state-of-the-art across several benchmarks.

Selective state space models (SSMs), such as Mamba, achieve strong per-token expressivity by making the time discretization step $\Delta$ a learned function of the input. However, in doing so, $\Delta$ ceases to represent a physical sampling interval, limiting its irregular time series modeling capability. Continuous-time SSMs, such as S5, preserve the physical meaning of $\Delta$ and handle irregular timestamps natively, but their dynamics remain linear time-invariant (LTI), limiting per-token expressivity. We propose \textbf{TIDES}, a selective SSM variant that reconciles selective and continuous architectures by moving input-dependence off the step size and onto the diagonal state matrix. As a result, $\Delta$ retains its physical meaning, tied to the state discretization, allowing the model to handle irregular timestamps natively without sacrificing the per-token expressivity that makes selective SSMs effective. We show this on a novel \emph{Fading Flash} experimental benchmark, a compact controlled diagnostic for sequence models that jointly tests input-dependence and extrapolation to out-of-distribution $\Delta$ values, and isolates the distinct failure modes of current state-of-the-art architectures that TIDES avoids by construction. On large-scale benchmarks, TIDES sets the new state-of-the-art average rank on UEA time-series classification and the Physiome-ODE regression benchmark.

Large Reasoning Models (LRMs) achieve impressive performance on complex reasoning tasks via Chain-of-Thought (CoT) reasoning, which enables them to generate intermediate thinking tokens before arriving at the final answer. However, LRMs often suffer from significant overthinking, spending excessive compute time even after the answer is generated early on. Prior work has identified the existence of an optimal reasoning length such that truncating reasoning at this point significantly shortens CoT outputs with virtually no change in performance. However, determining optimal CoT lengths for practical datasets is highly non-trivial as they are fully task and model-dependent. In this paper, we precisely address this and design Terminator, an early-exit strategy for LRMs at inference to mitigate overthinking. The central idea underpinning Terminator is that the first arrival of an LRM's final answer is often predictable, and we leverage these first answer positions to create a novel dataset of optimal reasoning lengths to train Terminator. Powered by this approach, Terminator achieves significant reductions in CoT lengths of 14%–55% on average across four challenging practical datasets: MATH-500, AIME 2025, HumanEval, and GPQA, while outperforming current state-of-the-art methods and reducing inference latency by more than 2$\times$ compared to the original LRM.

Recurrent models trained from scratch have recently become competitive on ARC-style reasoning tasks, but the usual framing around small recurrent backbones overlooks two important parts of the system: task-conditioned memory and synthetic augmentation data. We study this regime through CHARM, a compact hybrid ARC model that combines recurrent reasoning with structured task memory, synthetic data, and inference-time aggregation. In existing approaches, task-conditioned memory supplies a large hidden source of capacity, reaching more than 30x the size of the recurrent backbone. We introduce a compositional sparse embedding for task conditioning that reduces learned task-memory parameters by over 90% while improving pass@2 in controlled ARC-AGI-1 ablations. We also find that matched synthetic data is the largest measured driver of ARC-AGI-2 gains. For the recurrent backbone, recurrent depth helps only when balanced with learning horizon. Combining these ingredients, our system reaches 80.5% pass@2 on ARC-AGI-1 and 39.3% pass@2 on ARC-AGI-2 public evaluation.