NeurIPS 2026 Papers

Aligned language models refuse unsafe requests because RLHF widens the logit margin between refusal and affirmative tokens at the first decoded position. We call this scalar the refusal-affirmation logit gap and use it as a per-prompt diagnostic for alignment robustness. On three model families, alignment widens the gap on 97.5-99.8\% of toxic prompts, and gap closure tracks True ASR across suffix strategies (an internal consistency check, since our method optimises for gap closure). We present logit-gap steering, a gradient-free, forward-pass-only method that searches for in-distribution suffixes drawn from the model's own high-probability candidates whose cumulative effect closes the gap. Discovering 8 ensemble suffixes per family costs ${\approx}26{,}000$ forward-pass equivalents (${\approx}2$~min on one A100), ${\approx}125\times$ less than a single GCG search. Suffixes discovered on 0.5B-2B models transfer to 72B within family. An 8-suffix ensemble reaches 38--96\% True ASR across 13 models on AdvBench and HarmBench. Most suffixes have $10^{3}$-$10^{4}\times$ lower perplexity than GCG: under a published PPL-filter defense, our ensemble holds at 76.0\% (from 76.9\%) while GCG collapses from 64.7\% to 1.0\%.

In long-context use, large language models frequently synthesize answers from the meaning of a relevant context span rather than literally copy-pasting them. Identifying which attention heads perform this synthesis matters for interpreting long-context model behavior. Yet existing detectors miss these heads by construction: they reward heads whose attended token matches the generated token, a literal-copy criterion that captures where a head reads but not what it writes through its output-value (OV) circuit, the very mechanism that carries non-literal retrieval. We introduce Logit-Contribution Scoring (LOCOS), a write-aware detector that scores each head by the projection of its OV-circuit output onto the answer-token unembedding direction, contrasting needle and off-needle source positions in a single forward pass. Across three model families (Qwen3, Gemma-3, OLMo-3.1), mean-ablating the top LOCOS heads on the NoLiMa non-literal retrieval benchmark collapses ROUGE-L at lower head counts than every attention-based baseline; on Qwen3-8B, ablating 50 heads drives ROUGE-L from 0.401 to 0.000 while the strongest baseline still retains 0.292. The selected heads are retrieval-specific: parametric recall and arithmetic reasoning stay at baseline under the same ablation.

We study the $K$-armed logistic bandit problem, where at each round, the agent observes $K$ feature vectors associated with $K$ actions. Existing approaches that achieve a rate-optimal $\widetilde{\cal O}(\sqrt{dT})$ regret bound rely heavily on context diversity assumptions, such as strict positivity of the minimum eigenvalue of a context covariance matrix. These assumptions, however, impose strong restrictions on the context process, as they rule out the situation where the context vectors are concentrated in a low-dimensional subspace. In this paper, we propose SupSplitLog, which, to the best of our knowledge, is the first algorithm for logistic bandits that achieves $\widetilde{\cal O}(\sqrt{dT})$ regret without any context diversity assumption. The key idea is to split the collected samples into two disjoint subsets when constructing estimators; one is used to compute an initial-point estimator, while the other is used to apply a Newton-type one-step correction procedure. The splitting rule is carefully designed to balance the accuracy requirements of the initial-point estimator and the one-step correction procedure. Moreover, SupSplitLog strictly improves on the existing algorithms in terms of the dependence on dimension $d$ in the regret upper bound. Furthermore, SupSplitLog can be adapted simply to deduce a regret bound that grows with a data-dependent complexity measure, avoiding a direct dependence on $d$, which is favorable when the context vectors are concentrated in a low-dimensional subspace. We also provide experimental results that demonstrate numerically the superiority of our algorithm, validating the theoretical results.

We present an algorithm that distills transformer encoders into logical formulas. Interpretable surrogates for transformers are important for safety and scientific knowledge discovery. Recent work extracts sparse latent features from transformer activations, but these features alone do not express how information is composed across a sequence. Building on theoretical connections between transformer expressivity and linear temporal logic (LTL), we introduce iterated temporal decision trees (ITDTs), a model that can express any LTL formula as a sequence of decision trees. Each decision tree builds on the previous ones, producing LTL formulas of increasing temporal depth. Our algorithm distills a transformer encoder into an ITDT. Empirically, our algorithm recovers succinct and accurate formulas on established LTL learning benchmarks and distills interpretable rules for multiple protein sequence datasets from ESM-2-T6, an established protein language model.

Long-form technical text generation underpins knowledge-intensive workflows, yet remains challenging for large language models (LLMs) due to the need for globally consistent logical structuring and faithful technical reasoning beyond local coherence. Patent drafting is a canonical instance of this challenge, demanding holistic generation of a legally compliant and technically exhaustive document through sustained multi-expert collaboration. Existing approaches often focus on partial section generation or rely on manually crafted outlines, limiting scalable automation in realistic settings. In this work, we propose LogicTree-RAG, a logic tree-guided retrieval-augmented generation framework that induces a hierarchical logic tree as a global organizational backbone to organize and ground technical disclosures, without relying on expert-defined drafting priors. Each node in the logic tree represents a technical element and is constructed through evidence-guided recursive generation. A hybrid traversal mechanism then maps the logic tree into patent sections, enabling controllable and section-balanced generation. Extensive experiments show that LogicTree-RAG consistently improves content quality and language conformity over strong LLM-based baselines and achieves longer structured generation with high token efficiency, demonstrating the effectiveness of logic-centric generation for complex technical document drafting.

Accurately rendering temporal composition is fundamental to text-to-video (T2V) generation. While recent efforts incorporate temporal structure for long-horizon coherence, the intrinsic temporal understanding capability of T2V models remains underexplored. We observe that current models still struggle to respect canonical operators such as *before* and *after*, usually accompanied by missing entities and incorrect action binding. In this work, we propose LogicDirector, a test-time guidance framework that enforces temporal composition through executable logical specifications. By translating text prompts into first-order logic constraints over attention-derived evidence, LogicDirector constructs a neuro-symbolic verifier to evaluate noisy latent states during diffusion sampling for entity grounding, action binding, and temporal adherence. To enforce these constraints without expensive gradient-based optimization, we introduce a gradient-free Best-of-*n* latent search strategy, which preserves the model's native diffusion dynamics while steering logical satisfaction. To systematically study operator-level temporal composition, we curate TempBench, a diagnostic benchmark spanning four canonical relations, together with a hierarchical evaluation protocol that disentangles entity existence, event realization, and temporal correctness. Experiments on CogVideoX and Wan 2.1 demonstrate that our method significantly improves temporal composition while maintaining visual fidelity.

Transformers can perform in-context learning at depths far smaller than the number of optimizer updates suggested by existing gradient-descent interpretations. Prior work shows that an L-layer Transformer can emulate L gradient-descent updates, but this leaves open whether depth is merely a step counter. We show that it is not: in linear self-attention, each layer compounds the algebraic capacity built by previous layers, yielding a tight Θ(log k) depth characterization for computing the final result of k gradient-descent steps on diagonal quadratic tasks. Controlled experiments validate this mechanism: trained linear-attention models exhibit the predicted logarithmic depth transition, and kernel and oracle comparisons identify polynomial degree as the governing bottleneck. Broader experiments show the same depth signal under standard Transformer components, broader ICL regression families, and early-exit probes on Qwen2.5.

Thinking with images empowers vision-language models (VLMs) to use an external zoom-in tool for fine-grained visual search. Recent works employ supervised fine-tuning (SFT) and reinforcement learning (RL) with tool-specific designs to enable such capabilities. However, achieving logical reasoning alongside accurate and efficient tool use remains a significant challenge. To address this, we shift the focus away from increasing RL complexity and instead identify SFT data quality as a key bottleneck. We then propose LogT (short for **Log**ically **T**hink with Images), a fully automated SFT data engine that explicitly enforces logic-chain-guided reasoning. LogT generates an SFT dataset featuring challenging, visually grounded questions and structured tool-use traces with human-like logic. We validate our approach by fine-tuning two base models, Qwen2.5-VL-7B and InternVL3.5-8B, on the dataset and evaluating them across seven benchmark test sets. We show that LogT-SFT based on Qwen2.5-VL-7B outperforms the state-of-the-art trained with both SFT and RL, despite using 3.1$\times$ less data and $>$120$\times$ fewer GPU hours. Furthermore, applying a simple RL recipe without tool-specific design on LogT-SFT yields a model that surpasses prior methods with complex RL setups by a substantial 9.59-point margin. In addition to strong performance gains, our model exhibits higher tool-use accuracy and efficiency, as well as more structured and logically consistent reasoning traces, directly validating the effectiveness of our approach.

Time-series data are often sampled irregularly at high frequency and exhibit long-range dependencies, which makes long-horizon modelling difficult. Continuous-time models such as neural controlled differential equations (NCDEs) and neural rough differential equations (NRDEs) can handle irregular sampling, but they scale poorly to long sequences. Selective state-space models (SSMs) such as Mamba scale linearly with sequence length, but within a single block they provide limited recurrent mixing across hidden dimensions. We propose LogSig-SSM (Log-Signature Compression for State-Space Models), which first compresses long multivariate time series into a shorter sequence of tokens using multi-scale windowed log-signatures, and then processes these tokens with a selective SSM backbone. LogSig-SSM is scalable and robust to irregular sampling, combining log-signature tokens that capture higher-order cross-channel interactions with a selective SSM that models long-range dependencies. The model also admits a continuous-time interpretation as an NCDE/NRDE-style system driven by a log-signature-based input, in which selectivity induces an input-dependent rescaling of the latent dynamics. Across four benchmarks, namely long-sequence classification on UEA, high-frequency physiological regression on PPG-DaLiA, multivariate weather forecasting, and irregularly sampled clinical prediction on PhysioNet Sepsis, LogSig-SSM overall outperforms existing SSM and continuous-time baselines while significantly reducing training time and GPU memory usage.

Neural models such as YOLO and HuBERT detect local properties such as objects ("car") and emotions ("angry") in video frames and audio clips, with scores in [0, 1]. Lifting these scores to temporal properties over sequences enables applications such as query matching (e.g., "does the speaker eventually sound happy in this audio clip?") and ranked retrieval (e.g., "retrieve top 5 videos with a 10 second scene where a car is detected until a pedestrian is detected"). We formalize this problem of assigning Scores for TempOral Properties (STOPs) over sequences, given noisy score predictors for local properties. We propose LogSTOP, a scoring function that efficiently computes scores for temporal properties represented in Linear Temporal Logic. Empirically, LogSTOP with YOLO and HuBERT outperforms Large Vision / Audio Language Models by at least 16\% on query matching with temporal properties over objects-in-videos and emotions-in-speech, while matching or improving over temporal-logic baselines. On ranked retrieval with temporal properties over objects and actions in videos, LogSTOP with OWLv2 and SlowR50 improves mean average precision over zero-shot text-to-video retrieval baselines by 28\% and 18\% respectively.

The ability to reliably distinguish human-written text from that generated by large language models is of profound societal importance. The dominant approach to this problem exploits the likelihood hypothesis: that machine-generated text should appear more probable to a detector language model than human-written text. However, we demonstrate that the token-level signal distinguishing human and machine text is non-uniform across the hidden space of the detector model, and naively averaging likelihood-based token scores across regions with fundamentally different statistical structure, as most detectors do, causes a form of Simpson's paradox: a strong local signal is destroyed by inappropriate aggregation. To correct for this, we introduce a learned local calibration step grounded in Bayesian decision theory. Rather than aggregating raw token scores, we first learn lightweight predictors of the score distributions conditioned on position in hidden space, and aggregate calibrated log-likelihood ratios instead. This single intervention dramatically and consistently improves detection performance across all baseline detectors and all datasets we consider. For example, our calibrated variant of Fast-DetectGPT improves AUROC from $0.63$ to $0.85$ on GPT-5.4 text, and a locally-calibrated DMAP detector we introduce achieves state-of-the-art performance across the board. That said, our central contribution is not a new detector, but a precise diagnosis of a significant cause of under-performance of existing detectors and a principled, modular remedy compatible with any token-averaging pipeline. This will serve as a foundation for the community to build upon, with natural avenues including richer distributional models, improved calibration strategies, and principled ensembling with hidden-space geometry signals via the full Bayes-optimal decision rule.

We propose Log-Averaged Mirror Prox (LAMP), a linear-space primal-dual method for large-scale optimal transport. LAMP implements primal mirror prox updates by tracking an averaged dual sequence, reducing storage complexity from $\mathcal O(nm)$ to $\mathcal O(n+m)$ while preserving dense, GPU-friendly reductions. Consequently, LAMP preserves the last-iterate $\widetilde{\mathcal{O}}( nm\varepsilon^{-1})$ arithmetic complexity of conservatively parameterized primal-dual mirror prox. We further analyze LAMP as a direct optimal transport solver in a more performant parameter regime, providing a last-iterate sub-optimality certificate dependent on infeasibility and an explicit $\mathcal O(1/t)$ term. Moreover, we give a computable sufficient condition for best-iterate convergence to a saddle-point. Numerical experiments with an optimized CUDA implementation show that LAMP outperforms first-order baselines in several high-accuracy (entropic) optimal transport problems. LAMP is further shown to scale up to problems with $n=m=2^{18}$ marginal supports, which were previously beyond the reach of primal-dual first-order methods.

The quality of open-weight language models has dramatically improved in recent years. Sharing weights greatly facilitates model adoption by enabling their use across diverse hardware and software platforms. They also allow for more open research and testing, to the extent that users can use them as checkpoints, fine-tune them according to their needs, and potentially redistribute them. In some cases, however, concerns on modifying these weights towards unauthorized uses may outweigh the pros of giving users such a freedom. Defending against such adaptation is non-trivial: since an adaptive attacker can observe all weights and architectures by definition, they can reverse simple structural defenses, and use optimization to defeat the simplest locking mechanisms. In this work, we exploit the inference–training asymmetry of automatic differentiation as a novel defense axis. We propose DLR-Lock, a method where the purveyor of the model purposely replaces each pretrained MLP in their model with a deep low-rank residual network (DLR-Net) of comparable parameter count, forcing activation memory that grows linearly with depth during backpropagation. DLR-Nets are efficiently trained via module-wise distillation. We show that, beyond this memory overhead, DLR-Lock results in architectural mismatches that complicate the optimization landscape of standard fine-tuning, and a backward pass that incurs disproportionately more overhead than the forward pass. Our defense succeeds in withstanding adaptive attackers with full knowledge of the defense strategy while preserving the original model's capabilities. Experiments on LLM validate these claims.

Vision-Language Models (VLMs) are increasingly adopted as end-to-end driving planners, yet they suffer from cross-city domain shift. Existing methods treat this shift as a single quantity addressed by uniform alignment or single-site editing. We refute this view on two counts: ***(i) domain shift in VLM trajectory planners is inherently structured across modality and layer, and (ii) a locate-then-repair recipe matched to this structure outperforms every uniform-objective baseline***. We analyze domain shift via activation patching and uncover two regularities. **(F1) Modality axis.** Image and text tokens exhibit distinct shift patterns in each VLM. **(F2) Layer axis.** The layer-wise shift partition is architecture-determined and dataset-stable. These findings recast domain shift as a **two-dimensional tensor** indexed by *modality* (image vs. text tokens) and *layer*, motivating a simple principle: ***locate where shift occurs, and repair only there***. We instantiate this principle in **MoLaRx** (**Mo**dality–**La**yer **R**epair), a locate-then-repair framework. The Causal Importance Score (CIS) estimates the domain shift as a modality–layer tensor and partitions the layers into functional zones. CIS-Guided Selective Repair (CISR) then applies a matched operator (Maximum Mean Discrepancy (MMD) / low-rank adaptation (LoRA) / freezing) to repair each zone, with the composition adapting automatically to each architecture. We evaluate **MoLaRx** on three architecturally distinct VLMs (LFM2-VL-1.6B, Qwen2-VL-2B, and PaliGemma2-3B) across two cross-city benchmarks (nuScenes and Argoverse 2). MoLaRx attains the lowest cross-city trajectory-error gap on every (VLM, dataset) cell, while using $50$–$60\%$ fewer trainable parameters and $8.4$–$10.9$ ms lower inference latency. In contrast, the strongest scalar baseline (MMD-only) collapses on every cell—direct evidence that no scalar objective can repair structured shift. Code: [anonymous.4open.science/r/MoLaRx-E3B4](https://anonymous.4open.science/r/MoLaRx-E3B4/).

We study bandit convex optimization (BCO) when feedback is adversarially delayed and the comparator sequence is also adversarially non-stationary. This setting arises in online advertising, recommender systems, and adaptive dosing, but existing algorithms either absorb one of the two axes into a black-box global penalty or pay a worst-case cost that ignores the local interaction between delay and drift. We propose Oracle-DARS-DBCO, a delay-aware blocking algorithm whose block length $K$ and one-point smoothing radius $\delta$ satisfy the single identity $K\delta^{2}\asymp n^{2}$. This identity cancels the multiplicative coupling between delay deviation and one-point estimator variance, so the regret of each segment is paid only on that segment. We prove matching upper and lower bounds on the expected dynamic regret: $\widetilde{O}\bigl(\sqrt{n}\,m^{1/4}T^{3/4}+\sqrt{d_{\max}\,mT}\bigr)$ for convex losses and $\widetilde{O}\bigl(n^{2/3}m^{1/3}T^{2/3}+d_{\max}\,m\log(eT/m)\bigr)$ for $\alpha$-strongly convex losses, where $m=S_{T}+1$ is the number of stationary segments and $d_{\max}$ is the worst-case delay. The matching Rademacher-sign lower bounds $\Omega\bigl(\sqrt{d(S_{T}+1)\,T}\bigr)$ and $\Omega\bigl(d(S_{T}+1)\bigr)$ show that the delay price is segment-local. On fourteen scaling experiments, the fitted log--log exponents agree with the predicted ones within $0.04$, and Oracle-DARS-DBCO reduces regret by up to $9.17\times$ relative to the strongest non-restarting baseline on a piecewise-stationary benchmark.

Promptable segmentation foundation models such as SAM and its variants achieve strong performance with informative prompts, yet exhibit a striking prompt asymmetry where dense prompts like bounding boxes yield accurate segmentation, whereas sparse point prompts substantially degrade performance. This performance gap is widely observed, yet its internal decoder mechanism remains underexplored. In this work, we revisit the failure of point prompts from a causal perspective. By adapting activation patching to SAM-style decoders with a token-aligned intervention scheme, we localize the earliest recoverable divergence to a single early operation, i.e., the first decoder image-to-token cross-attention update, where prompt-conditioned information first enters the image stream. A sparse autoencoder further reveals that this mediation is concentrated in approximately 50 feature directions out of 4,096, and reverse corruption confirms that the identified layer is the entry point of a propagated image-stream route rather than a locally sufficient state. Guided by this localization, we propose the Causal Bottleneck Adapter (CBA), a lightweight residual module placed at the identified entry layer with the base model frozen. Evaluated on 97,996 medical image-target pairs across CT, MRI, ultrasound, and endoscopy, and validated across various SAM-family models, CBA closes 0.891 of the box-to-point gap with fewer than 9K trainable parameters, outperforming prior repair baselines up to 50× larger.

As large language models (LLMs) are increasingly integrated into high-stakes decision-making, the ability to reliably quantify uncertainty has become a critical requirement for safety and trust. However, current uncertainty quantification methods primarily operate at the output level, often failing to distinguish whether uncertainty arises from the model’s lack of knowledge or from ambiguity in the user’s input. While input-centric uncertainty quantification has recently emerged as a promising direction, it remains relatively underexplored and typically relies on coarse, input-level information. Consequently, users are provided with scalar uncertainty scores that offer little actionable guidance on which parts of the input should be clarified to improve reliability. To address this limitation, we propose Shapley-based input uncertainty Quantification (ShaQ), a framework for span-level attribution of input-induced uncertainty. Our approach models ambiguous spans in the input as players in a cooperative game and quantifies their contributions using Shapley values, defined via the weighted average of marginal reductions in conditional entropy obtained by clarifying each span coalition. Unlike existing input-level approaches, our formulation explicitly captures complex interactions among spans and provides a principled decomposition in which individual attributions sum exactly to the total input-induced uncertainty. We evaluate ShaQ on the AmbigQA and AmbiEnt benchmarks, where it achieves state-of-the-art performance in ambiguity detection. We further demonstrate its practical utility on a MediTOD benchmark, showing that ShaQ can precisely localize under-specified clinical utterances, facilitating more effective human-AI collaboration in high-stakes settings. Overall, our results show that ShaQ not only improves uncertainty estimation but also provides actionable insights for targeted input clarification. Codes are available https://anonymous.4open.science/r/ShaQ-0E39/README.md.

Understanding the internal routing of visual concepts is crucial for the safe and controllable adaptation of generative models. While concept localization has been widely studied in diffusion models, the emerging paradigm of next-scale visual autoregressive models remains largely unexplored. In this paper, we introduce LoCo, the first model-agnostic framework to localize where and when specific concepts emerge within autoregressive models. Driven by the native coarse-to-fine nature of next-scale generation, our method precisely maps conceptual knowledge across three distinct dimensions: Layer, Scale, and Position. To systematically localize and evaluate concept routing without context bias, we propose LoCoBench, a comprehensive dataset spanning 10 diverse categories. Extensive probing on image and video autoregressive models like Infinity, HunyuanImage-3.0, and InfinityStar shows that the localized positions are both interpretable and causally related to concept emergence. Building on these insights, we apply our localization method to three key applications: concept erasure, model personalization, and adversarial concept injection. Experiments demonstrate that our targeted intervention achieves state-of-the-art performance, substantially reducing computational overhead while preserving benign utility. Overall, our findings offer insights into how conceptual knowledge is routed during autoregressive generation, introducing a practical pathway for more interpretable, efficient, and secure adaptation.

Off-dynamics offline reinforcement learning (RL) aims to learn a policy for a target domain using limited target data and abundant source data collected under different transition dynamics. Existing methods typically address dynamics mismatch either globally over the state space or via pointwise data filtering; these approaches can miss localized cross-domain similarities or incur high computational cost. We propose Localized Dynamics-Aware Domain Adaptation (LoDADA), which exploits localized dynamics mismatch to better reuse source data. LoDADA clusters transitions from source and target datasets and estimates cluster-level dynamics discrepancy via domain discrimination. Source transitions from clusters with small discrepancy are retained, while those from clusters with large discrepancy are filtered out. This yields a fine-grained and scalable data selection strategy that avoids overly coarse global assumptions and expensive per-sample filtering. We provide theoretical insights and extensive experiments across environments with diverse global and local dynamics shifts. Results show that LoDADA consistently outperforms state-of-the-art off-dynamics offline RL methods by better leveraging localized distribution mismatch.