NeurIPS 2026 Papers

The strong zero-shot and long-context capabilities of recent Large Language Models (LLMs) have enabled highly effective list-wise reranking systems. Attention-based rerankers leverage transformer attention patterns as retrieval signals, but not all attention heads contribute equally: many introduce noise and redundancy, limiting both reranking quality and efficiency. In this work, we introduce CoRe heads, a small subset of retrieval-specialized heads identified through a contrastive scoring objective that rewards attention to relevant documents while penalizing attention to distractors. This contrastive criterion isolates highly discriminative retrieval heads and yields a strong training-free list-wise reranker. Our analysis reveals that retrieval-relevant computation in transformer rerankers is highly sparse and structurally localized: fewer than 1% of attention heads account for most reranking performance across models and datasets, and these heads consistently concentrate in middle transformer layers. This localization enables aggressive layer pruning for efficient reranking: removing the final 50% of transformer layers preserves nearly identical reranking accuracy while substantially reducing inference latency and GPU memory usage.

Prompt-level instructions are the main way users give language models preferences, safety boundaries, and content rules. But a model can remember a constraint and still violate it when generating text. This points to a gap between two abilities that are often treated as one: retrieving a constraint from context, and enforcing it during token selection. Across five model families (0.5B--32B parameters), we find a clear behavioral split: constraints that admit an alternative generation plan are largely followed, while token-level prohibitions fail whenever the prohibited continuation is locally dominant. Mechanistic analyses show that retrieval heads still attend to the constraint at the violation step, but the resulting signal either paradoxically enhances the prohibited token's logit or suppresses it too weakly to change the argmax. Decoding-time interventions can eliminate surface-form violations, but semantic compliance still requires higher-level verification. These results establish that constraint retrieval and constraint enforcement are separable capabilities, and that reliable token-level compliance requires explicit veto mechanisms beyond prompting.

The attention mechanism in Transformers links queries, keys, and values through a sequence of computations, while the principles that constrain its form remain unclear. This paper examines attention under natural requirements that formalize basic properties of attention computation as transformations between representation spaces, and asks which forms of attention computation are admissible. We show that similarity computation, score-to-weight mapping, and value aggregation are unified by dual structures induced by convex potential functions and their Legendre transforms. In particular, the standard inner-product score is recovered as a limiting case of a potential-based score, the score-to-weight map must take a gradient-map form under structural requirements on attention normalization, and the value-aggregation rule must take a dual-coordinate barycentric form under structural requirements on representative-point aggregation. These results show that the standard components of attention are structurally constrained by the underlying dual structure, while the choice of potentials provides explicit degrees of freedom for understanding and modifying attention mechanisms.

Dense video captioning describes multiple events in long videos. However, each event is predicted independently, ignoring the causal relationships with the other events in the video. These causal relationships are important for human-like descriptions of long videos. Our experiments show that exisiting vision-language models fail at predicting causal relationships in long-form videos. No prior work jointly generates captions and event timestamps and predicts causal relationships over multi-event sequences in long-form videos. We address this through two contributions. First, we create a benchmark dataset with 21.6K videos and 85K events by using a large language model to generate annotations from dense captions in YouCook2 and ActivityNet. We validate this benchmark dataset through multimodal alignment, vision-language models, and human judgment. Second, we propose a novel architecture for dense causal captioning that not only generates captions but also learns causal relationships. This novel architecture not only predicts causality effectively (0.73 F1 ActivityNet, 0.64 F1 YouCook2) but also improves the vision-language model's core capabilities: caption quality increases by +6.3 CIDEr on ActivityNet and +7.4 on YouCook2, while temporal grounding becomes accurate as causal predictions force precise event boundary localization.

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.

Dense correspondence matching has historically been bounded by simplifying spatio-temporal priors, such as smooth motion and rigid geometry. While effective for classical tasks, these foundations collapse at the frontier of Vision-Language-guided Image Editing and Generation (VL-IEG), where transformations yield correspondences that are perceptually obvious yet physically discontinuous. To transcend spatio-temporal priors, we draw inspiration from human cognition: learning a generalized model of identity-preserving visual consistency rather than relying on physical constraints. To achieve this, we introduce FreeMatching, a unified framework that extracts identity-preserving knowledge from foundation models and diverse datasets. The model's capability is forged through a two-stage training paradigm: supervised pre-training on diverse annotated datasets, followed by weakly-supervised refinement on data without dense annotations. Experimentally, FreeMatching not only achieves performance competitive with SOTA methods on classical benchmarks but also establishes the first strong baseline for the challenging VL-IEG task. Furthermore, we demonstrate its utility as a quantitative metric for evaluating identity-preserving consistency in generative editing, aligning closely with human judgment.

Modern retrieval-augmented systems typically access a corpus through a fixed similarity interface (either sparse or dense) that returns only the top-$k$ most similar documents for a query. While efficient, this design becomes a bottleneck for agentic search: exact lexical matching, conjunctions of weak clues, local context verification, and multi-step hypothesis refinement are difficult to realize with a conventional off-the-shelf retriever, and evidence filtered out early cannot be recovered downstream. This problem becomes particularly acute when agents must iteratively discover intermediate entities, combine partial evidence, and revise their search strategy. To address this problem, we study direct corpus interaction (DCI), a simple alternative in which an agent searches the raw corpus directly using general-purpose terminal tools (e.g., grep, file reads, shell commands, and lightweight scripts), without relying on embedding models, vector indexes, or retrieval APIs. This approach requires no offline indexing and adapts naturally to evolving local corpora. Across benchmarks for information retrieval (i.e., BRIGHT and BEIR) and end-to-end agentic search (i.e., BrowseComp-Plus), this simple setup substantially outperforms strong sparse, dense, and reranking baselines. It also attains strong accuracy on multi-hop question answering without relying on any conventional semantic retriever. Our results suggest that as language agents grow more capable, retrieval quality depends not only on reasoning ability, but also on how richly the corpus interface supports interaction. DCI therefore points to a broader interface design space for agentic search. Our code is available at https://anonymous.4open.science/r/dci-agent/.

Standard Retrieval Augmented Generation (RAG) is poorly matched to agent memory. Unlike large heterogeneous corpora, agent memory forms a bounded and coherent interaction stream in which many spans are highly correlated or near duplicates. As a result, flat top-$k$ similarity retrieval often returns redundant context, while summary-centric hierarchies can blur the subtle details that distinguish one candidate from another. We argue that agent memory should follow the principle of decoupling before aggregation: the system should first isolate reusable facts, updates, and distinguishing details from similar histories, and only then organise them for efficient retrieval. Based on this principle, we propose xMemory, which constructs a revisable hierarchical memory structure from original messages to segments, memory components, and groups. xMemory segments interaction history into local events, decouples each segment into memory components, aggregates related components into high-level groups using a sparsity--semantic faithfulness objective, and maintains this structure incrementally as memory evolves. At inference time, xMemory retrieves top-down, first selecting a compact backbone of complementary groups and components, and then expanding to segments and raw messages only when additional evidence reduces the reader's uncertainty. Experiments on LoCoMo and PerLTQA across diverse open source and closed source LLMs show consistent gains in answer quality and inference token efficiency, supported by analyses of redundancy, evidence density, and coverage.

Neural subgraph retrieval systems, which retrieve corpus graphs containing a query graph as a subgraph, are used in safety-critical applications such as drug discovery, hardware trojan detection, and molecular similarity search. Despite their importance, the adversarial robustness of these systems remains unstudied. We propose GRAP, the first adversarial attack framework targeting subgraph matching-based graph retrieval. \our jointly selects a budget-constrained subset of corpus graphs and computes per-graph edge perturbations to maximally degrade retrieval quality. We formalize two attack regimes: a ranking attack (with solver access) that corrupts pairwise relevance ordering while preserving relevance labels, and a solver-free top-K attack that directly displaces relevant results from retrieved sets. In both cases, we show that the resulting adversarial set functions are monotone and approximately submodular, enabling greedy subset selection with provable approximation guarantees. Experiments on five benchmark datasets against multiple state-of-the-art victim retrievers demonstrate that our method consistently and significantly outperforms all baselines in both gray-box and black-box settings, with and without solver access.

State-of-the-art neural network sparsification relies on alternating dense and sparse training phases. While computationally expensive, dense phases are critical for performance. Our empirical analysis reveals the mechanics behind this gain: upon reintroduction, previously masked parameters receive disproportionately large gradients that drive essential mask rewiring and parameter sign-flipping. Building on these insights, we optimize the performance-to-FLOPs Pareto frontier through three innovations: (1) zero-initializing masked parameters during dense phases to promote relevant sign flips, (2) truncating dense phases based on early mask convergence, and (3) replacing the dense phase entirely with a sparse rewiring phase that updates only the active mask and a small, high-gradient fraction of out-of-mask weights. Evaluated on CIFAR100 and ImageNet, our variant ARC (Alternating Rewiring Compression) matches baseline AC/DC performance at a fraction of the computational cost, establishing a superior performance-to-FLOPs frontier also compared to standard sparse-to-sparse algorithms.