NeurIPS 2026 Papers

Modern object detectors achieve strong performance on standard benchmarks, yet their robustness to contextual variation remains insufficiently investigated. Prior evaluations largely rely on aggregate metrics such as AP on uncontrolled distribution shifts, which conflate multiple factors and can obscure how performance degrades under context change. We introduce \textsc{ContextShift}, a controlled benchmark for evaluating context dependence by systematically manipulating object--context relationships while preserving object appearance. Built on COCO 2017, it isolates context as an independent variable both implicitly, through geometric transformations and explicitly via synthetic and natural background substitutions - with a compatibility axis based on normalized pointwise mutual information (NPMI) for continuous evaluation. Across diverse detector architectures, we observe a consistent degradation pattern: false negatives increase up to 227\% and prediction volume decreases down to -44\%, while false positives remain relatively stable or decline. This observed suppression behavior is not revealed by aggregate metrics such as AP, which can mask substantial recall loss and changes in prediction dynamics. Further analysis suggests that degradation is not primarily explained by reduced confidence, but is instead consistent with a reduced formation of valid detection candidates. Moreover, we find that performance along the statistical compatibility axis is non-monotonic, peaking at intermediate NPMI and degrading toward both extremes, indicating that statistical co-occurrence does not correlate linearly with effective visual context. Finally, we show that a training methodology based on context-aware augmentations rather than the dataset alone, which yields overall better models: every augmented variant outperforms the dataset-only baseline on both original (unmanipulated) test images as well as manipulated images, indicating that data-oriented augmentation can partially recover performance lost to prediction-suppression failures by exposing the model to object--context decoupling during training.

Transformers rely on a growing key–value (KV) cache to store context, causing linear memory growth and limited generalization beyond the training window. We show that contextual influence can instead be viewed as updates to a low-dimensional subspace in the model’s weight space, meaning context can be represented as a bounded, low-rank parameter update rather than token-level storage. Motivated by this, we propose PARADE, which internalizes context as parametric memory. It compresses past inputs into a fixed set of weight-space basis vectors and composes them via query-dependent routing into per-query weight updates, enabling streaming inference with constant memory and effectively unbounded context. Experiments show that PARADE matches full attention on long-context tasks and outperforms efficient baselines, especially when relevant information lies far beyond the attention window, while improving as context exceeds the model’s training length.

Biological recurrent circuits can reuse a common substrate by changing its operating regime rather than replacing the circuit itself. This principle is difficult to isolate in fully adaptive neural models, where recurrence, controllers, readouts, and internal representations can all absorb task structure. As a result, successful multi-context behavior does not reveal which route made the computation possible. We introduce Constrained Modulatory Reservoirs (CMR), a fixed-substrate framework for studying such reconfiguration with the main bypasses removed. The primary dynamics and low-rank perturbation basis are fixed, and the readout is denied direct access to the cue, forcing input-dependent adaptation to act through a limited set of recurrent connectivity changes. In transient-cue channel-selection tasks, this restriction makes a capacity transition visible. Low-loss behavior appears only when the available perturbation directions are sufficient, the selected change is retained across the cue-to-signal gap, and the primary dynamics express it during the response window. Controls show that this effect is not explained by low-dimensional cue representation alone. Allowing the substrate to learn hides the bottleneck, while activity- and gain-level alternatives fail to reproduce the same regime. CMR therefore reframes flexible recurrent computation as a question of where adaptive influence is allowed to enter a fixed dynamical system, making capacity, memory, and expression experimentally separable.

Chunked prefill has become a widely adopted serving strategy for long-context large language models, but efficient attention computation in this regime remains challenging. Existing sparse attention methods are primarily designed for one-shot prefill and do not translate efficiently to chunked prefill: block-sparse kernels lose efficiency when the query length is limited by the chunk size, while fine-grained pattern search becomes costly when repeated over the accumulated KV cache at every chunk. QUOKA, a recent method that directly targets chunked prefill, avoids sparse-kernel overhead but relies on query-subsampled, token-level KV selection, which can miss query-specific KV entries and introduce explicit KV-copy overhead. To address these limitations, we propose CompactAttention, a chunked-prefill attention mechanism based on Block-Union KV Selection. CompactAttention treats 2D block-sparse masks as KV-selection signals rather than direct sparse-kernel execution plans, and converts them into GQA-aware per-group KV block tables through Q-block union and intra-group union. This construction produces the minimal block tables that preserve all KV blocks selected by the input masks under paged execution constraints, enabling selected KV blocks to be accessed in place without explicit KV compaction. On LLaMA-3.1-8B-Instruct, CompactAttention maintains accuracy close to dense attention on the RULER benchmark while delivering up to 2.72$\times$ attention speedup at 128K context length under chunked prefill.

Scaling Transformers to long contexts is constrained by the quadratic cost of self-attention and the linear growth of key-value cache memory transfer. Sparse attention mitigates this by retrieving only relevant tokens, but current approaches either require large-scale training or, within the training-free regime, rely on semantically coarse heuristics or expensive clustering that is difficult to update efficiently during decoding. We introduce CommunityKV, a framework that formulates sparse attention as a community detection problem. CommunityKV constructs a token graph from the $QK^T$ scores already computed during standard prefill, and partitions the graph into communities to enable retrieval of semantically coherent token groups. A local update rule assigns newly generated tokens to communities in constant time, enabling sparse retrieval throughout streaming decoding without global re-partitioning. We evaluate CommunityKV on open weight models and long context benchmarks, demonstrating that our approach achieves up to $1.86\times$ higher decoding throughput than exact FlashAttention-2 with negligible accuracy degradation.

Reliable deployment of time series classifiers in high-stakes domains requires explanations that capture the interplay between localized patterns and global temporal context. However, existing post-hoc methods often treat local features in isolation, failing to account for long-range dependencies. We propose CoLaX, a context-aware explanation framework that explicitly embeds local dynamics within a global temporal structure. By employing a principled context-conditioning mechanism, CoLaX models how global context modulates the importance of local patterns, revealing not only what is important but also when and why it matters. Extensive benchmarks demonstrate that CoLaX yields significantly more faithful and stable explanations than state-of-the-art baselines, demonstrating that joint local-global modeling is essential for reliable time-series interpretability.

We present CivBench, an open-source benchmark for evaluating language model agents in long-horizon, tool-mediated environments through the Model Context Protocol (MCP). A single episode spans 300+ turns and produces thousands of tool calls over a large action space, requiring sustained planning, state monitoring, and execution under partial observability. The environment exposes 76 MCP tools and a narration layer that converts visual game state into structured text. We use CivBench to characterise agent behaviour across four model families in 23 admissible runs. The sample is a pilot, not a model ranking: aggregate outcomes do not reliably discriminate models at this scale. Instead, we introduce two interface-level metrics that the environment makes measurable: Proactive Monitoring Rate (PMR), capturing whether agents actively query latent strategic state, and RAG@10, capturing whether commitments stated in structured planning reflections are executed within ten subsequent turns. Across runs we observe two consistent patterns under a shared playbook protocol. Agents under-monitor strategically relevant state that is available but requires explicit querying: despite playbook guidance to query victory progress every 20 turns, agents do so only every 30–75 turns, and in 7 of 20 detectable defeats they failed to query within the 20-turn warning window before game end. Agents also frequently fail to execute near-term commitments stated in their own planning reflections (RAG@10 between 48% and 66% across models). Both patterns arise despite tool access and explicit guidance, and we interpret them as deviations under instruction rather than absences of capability. We release the environment, scenarios, logs, metrics, and analysis pipeline at https://anonymous.4open.science/r/civbench/README.md.

Large language models (LLMs) rely on long-horizon tool invocation sequences for complex tasks, where each invocation can alter the task state and condition subsequent decisions. In long-horizon tool use, final-outcome rewards provide weak credit assignment over long interaction traces. Step-level rewards can offer more targeted feedback, but obtaining reliable step supervision often requires human or LLM judgment, or additional rollouts to estimate the downstream effect of an intermediate decision. In this paper, we argue that effective tool-use agents should estimate the long-horizon value of a possible next tool invocation before executing it. This objective requires comparative supervision over alternative invocations under the same context, while logged trajectories only contain the invocation that was actually taken. Therefore, we propose Comparative Inference for Tool-use Agents (CITA). CITA trains a Comparative Inference Model (CIM) from paired signals that combine observed tool behavior, scalable supervision from a Bayesian tool-graph simulator, and semantic judgments from LLM-based comparison. The resulting CIM learns to estimate how likely a possible next tool invocation is to support final task success under the current context. Across three tool-use benchmarks and multiple backbone LLMs, CITA consistently improves Tool F1 and task success. Additional analysis shows that CIM learns accurate step-level value estimates for comparative tool choices.

Language models (LMs) raise an alluring alternative to vector-based retrieval: \emph{generating} a relevant answer, entirely based on an in-context corpus. In spite of this appeal, prior studies consider proprietary systems or the smaller-scale reranking task, leaving corpus-scale in-context retrieval largely unexplored. In this work, we present the first systematic study of in-context retrieval on two scales practical retrievers demand: \emph{million-token} corpora and \emph{length-generalization} far beyond training-time sizes. We first introduce BlockSearch, a 0.6B LCLM retriever whose architectural and training modifications improve over prior LM baselines and length-generalize up to $10\times$ beyond its training regime. However, its retrieval still collapses under more extreme extrapolation. We trace this failure to an \emph{attention dilution} effect: as the corpus grows, irrelevant documents dominate the softmax denominator, minimizing the normalized mass on the gold document even as its pre-softmax score stays high. Motivated by this analysis, we introduce \emph{length-aware} adjustments to the attention softmax and \emph{document-level sparse attention}, improving retrieval at million-token scale, better performance than the concurrent $7\times$-larger MSA system, and comparable performance to dense retrieval. Together, our results position in-context retrieval as a viable alternative to classical retrieval, emphasizing attention control under extreme context growth as a new challenge.

Effective machine learning depends not only on how we model data, but also on what data we choose to collect. While large sequence models have revolutionized data modeling, the problem of automated data selection, or "intrinsic curiosity", remains a significant challenge. Classic approaches incentivize exploration by rewarding an agent based on its "learning progress", which measures how much a newly acquired observation improves a world model's predictive ability. However, evaluating these rewards traditionally requires expensive inner loops of gradient descent updates within each trajectory, rendering them computationally impractical at scale. In this work, we investigate whether the emergent in-context learning (ICL) capabilities of sequence models can eliminate this bottleneck by serving as immediate, update-free world models. Specifically, we evaluate whether an exploration policy can be trained to maximize learning progress, using solely the prediction errors and counterfactual context manipulations of an in-context learner. We first prove that in general Markov decision processes, this is in fact impossible in an unbiased way: the resulting intrinsic rewards either suffer from nuisance terms that bias their estimation of true learning progress, or they cannot be implemented using an in-context learner's prediction errors. Conversely, we prove a positive result for a broad subclass of non-temporal settings, encompassing active learning and Bayesian Experimental Design: here, ICL-derived rewards successfully bound and asymptotically converge to the true learning progress. We corroborate our theory with controlled experiments across continuous and symbolic environments, demonstrating that our ICL-driven framework successfully trains curious data-collection policies that explore optimally.

Prior-fitted networks (PFNs) are a promising class of tabular foundation models that perform in-context learning, whereby the entire labelled training set is supplied as context, and predictions for test queries are produced in a single forward pass. However, the quadratically scaling self-attention mechanism in many PFN architectures makes inference prohibitive for very large training datasets. We propose CRUMB (Clustered Retrieval Using Minimised-MMD Batching), a three-stage inference wrapper that (i) clusters the test queries, (ii) selects a small, distributionally matched training subset for each cluster by greedily minimising the maximum mean discrepancy (MMD), and (iii) runs exact PFN inference on each reduced-context batch. CRUMB is architecture-agnostic and requires no retraining. On the 51-dataset TabArena benchmark, evaluated across three PFN architectures (TabPFNv2, TabICLv1, TabICLv2), we show that CRUMB outperforms similar state-of-the-art context selection strategies. We also show that CRUMB is resilient to covariate drift, as the MMD-minimisation step naturally helps align the training context distribution to match the current test batch distributions.

Feedforward foundation models have recently shown remarkable 3D reconstruction capabilities. However, existing models exhibit large tracking drift in long-context streaming reconstruction due to error accumulation. In this paper, we revisit loop closure with streaming reconstruction foundation models to enable accurate, drift-free, kilometer-scale reconstruction. Specifically, our method detects loop candidates through global descriptor retrieval, and constructs loop-conditioned windows to estimate the relative poses between looped frames. Given the observation that our adopted streaming reconstruction backbone produces a globally consistent scale, we optimize all frame poses on the SE(3) manifold with sequential and loop closure constraints, avoiding the pose graph optimization on the Sim(3) or higher-dimensional SL(4) manifolds employed in prior works. Extensive experiments show that our method reduces drift and produces consistent geometry on kilometer-scale sequences, significantly outperforming the state of the art. The code will be made public.

Language model agents are becoming proficient executors at isolated, short-horizon tasks such as software engineering and customer service. Yet real-world challenges require a combination of sophisticated skills that remain largely untested in agents: (1) navigating long horizons amid uncertainty; (2) acquiring information in noisy environments; (3) adapting to a changing world; (4) orchestrating multiple moving parts toward a coherent goal. We introduce CEO-Bench, which evaluates these capabilities together by simulating a representative real-world task: operating a startup for 500 days. Given diverse and realistic company management tools, business databases, and social media, an agent needs to design pricing strategies, allocate operating budgets, analyze business data, respond to unexpected competitor moves, and more. Our evaluation shows that most state-of-the-art models struggle to succeed in this environment, and only one model (GPT-5.5) finishes the simulation above its $1M starting balance. CEO-Bench reimagines the role of AI in the future, shifting from solving isolated tasks to driving sustained, adaptive progress over time. We open source full code and trajectory.

Reinforcement learning (RL) is essential for adaptive decision-making, e.g., robust Meta-RL and domain randomization (DR), and efficient post-training of large language models (LLMs), where tasks are commonly structured as Markov decision processes or prompts. A central bottleneck in all these settings is the cost of agent-environment interactions required for task selection and policy optimization. Model predictive task sampling (MPTS) has emerged as a promising approach to improve sample efficiency by querying informative tasks via a risk predictive model (RPM) trained on optimization history. However, existing RPMs suffer from two failure modes: (i) sparse optimization histories, which limit RPMs' generalization across the task space, and (ii) a one-step lag in risk approximation, which causes biased difficulty estimation as the policy evolves. This work formalizes these failure modes, derives lagged-risk decompositions that expose the resulting temporal bias, and recasts risk prediction as an in-context sequence modeling problem. The developed RPM integrates a one-step look-ahead mechanism with a temporal-weighted sliding window to mitigate data non-stationarity and scarcity, without requiring additional rollouts. Empirically, our approach yields more reliable difficulty estimates and consistent performance gains across robust Meta-RL, DR, and prompt curriculum RL post-training of LLMs.

LLM-based search agents are widely used for information-seeking tasks, but their reliance on external tool returns introduces a critical security risk: web content retrieved during execution is untrusted, exposing agents to prompt injection and goal hijacking. Prior work on search-agent safety primarily focuses on static web-content injection, but modern agents issue follow-up queries and cross-check competing sources, so a single injected page is often diluted or rejected. We show that the channel delivering search and page observations is a fragile security boundary: beyond exposing the agent to a single poisoned page, a mediated search interface can repeatedly steer how the agent gathers evidence and forms its final answer. Under a constrained tool-intermediary threat model, appending only one controlled result per query can substantially increase attack success when the evidence is coordinated across the agent's trajectory. We study this setting with a strategy-driven long-horizon attack system and introduce Authority-Chain Hijack (ACH), an expert-refined strategy that turns isolated search-result and page-content manipulations into a coherent evidence chain across seemingly corroborating sources. ACH achieves the highest Overall ASR among all baselines, reaching 56.0\%/85.0\% ASR/MaxN\,ASR on the SafeSearch benchmark. We further introduce Trace-Guided Strategy Evolution (TGSE), which automatically improves reusable attacker strategies from execution traces, replacing manual redesign with trace-driven refinement and further raising these to 71.4\%/95.0\%.

We study long-context balanced entropic optimal transport (OT) attention on TPU hardware through a stopped-base, fixed-depth tail-refinement surrogate. After a stopped $T$-step Sinkhorn solve, we unroll a short refinement tail and differentiate that surrogate exactly. For the production $R=2$ case, the backward pass contains four staircase plan factors. We prove an exact one-reference-tile schedule: the $R=2$ score cotangent is a single reference plan tile times an explicit modifier field built from vector cotangents and dual differences. This yields block-wise cost $O((T+R)LW)$, $O(Ld)$ input storage, and $O(L)$ additional HBM usage for fixed head dimension $d$ and band width $W$. We also formalize the current \texttt{dustbin\_block} path as the same balanced surrogate on an augmented support, so the schedule lifts to the gap-aware transport path used in our TPU runs. We provide a local surrogate-bias bound, an a posteriori bias certificate, and a projective contraction certificate for strictly positive active blocks. On synthetic masked problems, the optimized kernel matches exact autodiff of the same centered surrogate to within $10^{-5}$--$10^{-10}$. On TPU v6e-8, a four-configuration Pfam screen completes end-to-end, and a promoted balanced $R=2$ run sustains roughly $8.5$ examples per second through a three-hour budget, reaching step $1437$. Held-out Pfam test shards improve reconstruction from $3.17$ to $0.99$ and sparse CE from $5.86$ to $5.69$ relative to step $0$. These results support exact fixed-depth backward theory, a theorem-matching gap-aware bridge, and trainability evidence for the production path.

Modern large language models increasingly require long contexts for reasoning and multi-document tasks, but attention's quadratic complexity creates a severe computational bottleneck. We present *Block Sparse Flash Attention* (BSFA), a drop-in replacement that accelerates long-context inference while preserving model quality. Unlike methods that predict importance before computing scores, BSFA computes exact query-key similarities to select the top-$k$ most important value blocks for each query. By comparing per-block maximum scores against calibrated thresholds, we skip approximately 50% of the computation and memory transfers for pruned blocks. Our training-free approach requires only a one-time threshold calibration on a small dataset to learn the per-layer and per-head attention score distributions. We provide a CUDA kernel implementation that can be used as a drop-in replacement for FlashAttention. On Llama-3.1-8B, BSFA achieves up to $1.13\times$ end-to-end speedup on LongBench with only a $1.1\%$ accuracy drop, and up to $1.24\times$ on Needle-in-a-Haystack retrieval at a $1\%$ accuracy drop. The attention kernel itself accelerates by up to $1.38\times$. We compare BSFA against five recent sparse attention baselines (SpargeAttention, MInference, FlexPrefill, XAttention, and BLASST), and verify the method on Qwen2.5-7B and on A6000 and H100 GPUs. The implementation is available at https://github.com/Anonymous44414/Block-Sparse-Flash-Attention.

Long-horizon agents increasingly operate across many steps, tools, and observations. In this setting, the relevant oversight question is not only whether each action is locally valid, but whether the evolving trajectory still corresponds to the task the user authorized. Drift can accumulate quietly: an agent may call the right tool with plausible arguments at every step, while its prefix moves toward a broader role, an adjacent objective, or evidence the user never supplied. Existing monitors mostly check local compliance, deliver final-trace verdicts, or score generic risk; they do not directly estimate this prefix-level relation. We introduce ontological trust, a task-conditioned property of trajectory prefixes, and instantiate it as RGE, an online monitor that decomposes trust along Role, Goal, and Evidence. RGE uses LLMs only to derive structured representations. State updates, projections, and intervention decisions are deterministic, so the output is a replayable and auditable trust trajectory rather than a single end-to-end judge verdict. We construct a cross-domain trajectory corpus from OSWorld, FinanceBench, and EICU-AC, covering benign executions, prefix-paired drift, and pseudo-consistency failures. On this corpus, RGE outperforms adapted rule-, judge-, and shield-style baselines on prefix-paired drift detection. With the two larger estimator models, it exceeds 93\% Drift F1 on every benchmark while keeping benign coverage at or above 95.8\%. Pseudo-consistency is harder: detection depends on whether task completion is externally visible, a structural limit we characterize empirically.

We study minimal attention-only transformers under all-token corruption and show they admit a two-stage empirical Bayes interpretation. A single attention step computes a kernel-weighted posterior mean with respect to the empirical distribution defined by the context. Depth refines this distribution through particle dynamics (Stage 1), while a long-range skip-connection carries the noisy input as a query for posterior inference (Stage 2), revealing distinct statistical roles for depth and attention residuals. The framework isolates a minimal setting in which the context itself induces a depth-dependent energy landscape governing in-context inference. We show that effective denoising can emerge without an explicit noise schedule: a fixed kernel bandwidth and finite integration horizon suffice, yielding a principled depth–noise relationship. We further establish a posterior-mean recovery guarantee for a class of well-behaved priors, where the empirical estimator converges to the Bayes-optimal predictor under asymptotic conditions. Connecting these dynamics to reverse-diffusion limits, our results provide a statistical interpretation of attention as in-context inference via sample-based posterior estimation, without explicit density modeling.