NeurIPS 2026 论文列表

Transformers acquire in-context learning abilities through abrupt phases during training, often unfolding over multiple stages in which key circuits, such as induction heads, emerge. In this work, we characterize the dynamics underlying the emergence of such circuits across these stages. We focus on a synthetic associative recall task, where sequences are drawn from random maps between a permutation group and a vocabulary, and the model is required to complete the mapping of a permutation by retrieving the corresponding association from context. For this task, we study the gradient-flow trajectories of a simplified two-layer attention-only transformer. Leveraging symmetries in both the transformer architecture and the data distribution, we derive *closed-form parameter dynamics* for both attention layers. We identify a conservation law that couples parameters across layers and governs their joint learning. This law reveals how initialization controls the *timescales* over which such circuits emerge. In the vanishing-initialization limit, we characterize the gradient-flow trajectory, showing how training jumps from saddle to saddle. Finally, we provide empirical evidence across different architectural choices, validating our simplifications and extending the insights from our analysis beyond the simplified setting.

We extend the Bayesian learning rule by explicitly incorporating the compositional structure of neural networks. The central observation is that backpropagation and exponential family variational inference share a common Lagrangian duality structure. Leveraging this connection, we develop a new Bayesian learning framework in which adjoint variables backpropagate layerwise sensitivity signals that converts into loss-site natural parameters. For Gaussian weight posteriors and Gaussian layer-state projections, the framework recovers deterministic back-propagation and Hessian backpropagation via delta-method approximations, while providing a unified perspective on localization-style prediction rules and moment matching methods. We also derive a new variant of IVON, clarify how its gradient and Hessian estimates differ from those of the standard version, and demonstrate how neural network structure can be incorporated into Bayesian learning in practice.

Sparse dictionary learning (and, in particular, sparse autoencoders) attempts to learn a set of human-understandable concepts that can explain variation on an abstract space. A basic limitation of this approach is that it neither exploits nor represents the semantic relationships between the learned concepts. In this paper, we introduce a modified SAE architecture that explicitly models a semantic hierarchy of concepts. Application of this architecture to the internal representations of large language models shows both that semantic hierarchy can be learned, and that doing so improves both reconstruction and interpretability. Additionally, the architecture leads to significant improvements in computational efficiency.

Clinical case search aims to retrieve disease-consistent prior cases from electronic health record databases for a natural-language patient query under a retrospective retrieval objective. The task is challenging because patient queries contain heterogeneous evidence, including symptoms, diagnoses, laboratory results, treatments, and timelines. These fields are often only partially specified, and fixed retrieval pipelines cannot adapt from prior search trajectories. Existing clinical retrieval methods typically flatten the query into a single representation, use static retrieval pipelines, or lack an explicit mechanism for cross-case adaptation. \textbf{PatientSearch-SE} addresses retrospective disease-centric clinical case search with three components: a \emph{sufficiency-aware planner} that separates dimensions suitable for direct matching from those requiring hypothesis completion, a \emph{hierarchical tool orchestration} module that maps these states to retrieval and re-ranking actions, and a \emph{trajectory-guided memory consolidation} mechanism that writes reusable disease cards and strategy templates without parameter updates. We also propose an evaluation framework that combines disease-centric retrieval outcomes with trajectory-level process probes for retrieval behavior analysis. Experiments on PMC-Patients and MIMIC-IV show higher disease-centric retrieval metrics for PatientSearch-SE than for the compared same-interface sparse, dense, and reasoning-based baselines, with larger gains on PMC-Patients and more modest gains on MIMIC-IV. Its memory-accumulation results are consistent with controlled external-memory adaptation, and its retrieved cases provide additional evidence for diagnosis-masked downstream few-shot RAG prediction under this retrospective protocol.

Scaling Graph Neural Network (GNN) training to large graphs commonly relies on mini-batch sampling, but incomplete neighborhoods can degrade approximation quality and training stability. Memory-enhanced methods mitigate this issue by reusing historical node embeddings to approximate full-neighborhood aggregation; however, they reconstruct pseudo full-neighborhood messages inside every training iteration, repeatedly fetching out-of-batch memories and aggregating over large neighborhoods. In this work, we identify this under-examined inner-loop bottleneck and propose IncAgg, an efficient memory-enhanced GNN training framework based on incremental aggregation. IncAgg periodically pre-aggregates historical embeddings into a global memory baseline, and during mini-batch training propagates only the current in-batch residual before fusing it with the stored global context. This design preserves historical full-neighborhood information at a configurable refresh frequency while avoiding per-iteration out-of-batch memory fetching and aggregation, without modifying the underlying GNN architecture. We provide theoretical analyses showing that IncAgg reduces gradient-carrying aggregation and memory-access costs while retaining the residual-based stability principle of memory-enhanced estimators. Extensive experiments on four large-scale benchmarks and four representative GNN backbones show that IncAgg matches or improves the accuracy of strong memory-enhanced baselines while achieving substantial training acceleration, with up to $18.6\times$ speedup in individual settings and up to $13.8\times$ geometric-mean speedup on large graphs.

For LLM agents to be useful beyond simple chat, they must hold sensitive user context such as calendars, credentials, health records, and financial data. We study whether the mere presence of such secrets in a model's context window introduces hidden correlations into the model's benign outputs, allowing reconstruction even when the model correctly refuses direct extraction. We further study whether an adversary can actively engineer prompts that amplify this effect, using the model as a covert carrier to transmit secrets through seemingly innocuous text. In both cases, this limited leakage is exploited using a novel _adaptive_ attack that assumes black-box access to the underlying model. In controlled experiments across eight proprietary models, we find that 2-digit in-context secrets are reconstructed with near-perfect accuracy and 4-digit secrets at 82% exact match, all from outputs the model produces in response to ordinary, non-adversarial requests. More capable models leak more: stronger instruction-following amplifies sensitivity to in-context secrets, making leakage a byproduct of capability rather than a bug to be patched. We show this leakage enables two practical attacks: (1) a trained classifier that infers semantic predicates about user memories (e.g., health conditions, financial events) from routine natural-language outputs, and (2) an RL-trained adversary that extracts full Social Security Numbers from a production-style agent at 100% success.

Continual learning seeks to acquire new knowledge while preserving previously learned representations. Despite the success of prompt-based methods, they remain fragile in domain-incremental learning scenarios, where shifts in the input distribution cause representation drift in shared layers and lead to forgetting. We introduce InTAct, a method that preserves functional behavior in shared representations without freezing parameters or retaining past data. InTAct constrains updates within activation regions associated with prior domains, while allowing flexible adaptation elsewhere, thereby stabilizing neuron functionality rather than directly restricting parameter values. The approach is architecture-agnostic, integrates seamlessly with existing prompt-based frameworks, and matches or outperforms state-of-the-art methods on domain-incremental benchmarks.

Key-value (KV) caches are essential for efficient Large Language Model (LLM) inference, but their memory footprint grows linearly with context length and batch size. Low-bit KV-cache quantization reduces this footprint, yet uniform quantization is vulnerable to high-magnitude outlier channels, while outlier-aware mixed-precision methods often introduce extra buffers, indexing, or channel permutation that weakens their system-level benefit. This paper presents InQuant, an in-place mixed-precision KV-cache quantization method that preserves salient outlier channels without changing the physical 4-bit packed layout. The key observation is that selected low-saliency channels near outliers can tolerate small approximation error; InQuant therefore reuses their storage slots to hold the extra 4-bit nibble needed by 8-bit outlier values. To make this layout practical, InQuant combines sampling-based channel saliency estimation, saliency-aware neighbor-slot reuse, stride-based handling for adjacent outlier groups, and descriptor-guided marker recovery during dequantization. Across six LLMs and long-context workloads, InQuant reaches a fixed 4.0$\times$ physical KV-cache compression ratio, preserves accuracy close to strong mixed-precision baselines, and reduces quantization/dequantization latency by 2.2$\times$--3.4$\times$ compared with representative outlier-aware methods.

Industrial anomaly detection benchmarks have advanced rapidly, yet they remain largely static and therefore underrepresent deployment settings in which inspection systems must be extended from one industrial component to another over time. In our setting, continual industrial anomaly detection is framed as a task-incremental problem: each component defines a task, while the learning objective remains semi-supervised. The goal extends beyond conventional anomaly detection, since models should preserve their performance on all components as they are sequentially exposed to new components. This framing is especially relevant because many existing industrial pipelines either assume fully supervised defect labels or study only isolated datasets, leaving open the question of how semi-supervised anomaly detectors behave across a sequence of heterogeneous components. We address this gap by organizing the benchmark as a sequence of related component-level tasks and by defining multiple scenarios, including random, easy-to-hard, and hard-to-easy. In addition to single-dataset continual scenarios, we propose challenging scenarios spanning across multiple datasets. We further evaluate continual performance from two complementary perspectives: image-level anomaly detection and pixel-level anomaly localization. By recasting industrial anomaly detection as a continual benchmark rather than a static train/test problem, our work establishes a concrete path toward studying knowledge retention, adaptation, and generalization in evolving industrial environments.

Critical transitions – abrupt, often irreversible changes in system dynamics – arise across human and natural systems, often with catastrophic consequences. Real-world observations of such shifts remain scarce, preventing the development of reliable early warning systems. Conventional statistical and spectral indicators, such as increasing variance, tend to fail under realistic conditions of limited data and correlated noise, whereas existing deep learning classifiers do not extrapolate beyond their training data distribution. In this work, we introduce TipPFN, an in-context learning (ICL) framework that uses a prior-data fitted network to infer a system's proximity to a critical transition. Trained on our novel synthetic data generator, which is based on canonical bifurcation scenarios coupled to diverse, randomized stochastic dynamics, TipPFN flexibly capitalizes on contexts of various sizes, complexity and dimensionalities. We demonstrate robust, state-of-the-art early detection of critical transitions in previously unseen tipping regimes, sim-to-real examples, and real-world observations in both ICL and zero-shot settings.

Many stochastic gradient methods are believed not to converge when the noise in stochastic gradients has only a finite $p$-th moment for $p\in\left(1,2\right)$, a setting known as the heavy-tailed noise assumption. However, some recent studies have found that Stochastic Gradient Descent ($\textsf{SGD}$), without any modifications to its update rule, can surprisingly converge in expectation for convex problems with bounded domains, highlighting the potential of classical stochastic gradient methods. Inspired by this recent progress, we provide a comprehensive study of stochastic optimization under heavy-tailed noise and establish new in-expectation convergence results for Stochastic Mirror Descent ($\textsf{SMD}$) in convex optimization and $\textsf{SGD}$ in nonconvex optimization. Notably, our results not only hold without algorithmic changes but also avoid restrictive assumptions such as bounded domains that were imposed in prior work. More importantly, our analysis provides a new, elegant, yet powerful framework for studying heavy-tailed stochastic optimization, opening a new route to understanding first-order stochastic gradient methods.

In-context learning has recently been linked to implicit gradient descent in linear self-attention models, suggesting that context can induce a forward-pass update. Retrieval-augmented generation (RAG) also relies on context, but retrieved documents are usually treated as static evidence rather than signals for adaptation. We study RAG as an in-context optimization process. First, we show that one linear self-attention layer can implement one gradient-descent step on a unified linearized RAG objective covering both projection-based and dot-product retrieval interfaces. This gives an exact regime where retrieval-augmented prediction and in-context optimization coincide. We use this result not as a literal model of LLM computation, but as a guide for adapting the interaction between queries and retrieved evidence. We then test the boundary of this correspondence: it remains stable under controlled linear extensions, but becomes feature-distribution dependent under nonlinear architectures. Finally, we turn this view into a lightweight method for frozen RAG LLMs. The method keeps the retriever and backbone fixed, and predicts a context-conditioned update to a generator-side evidence-use interface. Across seven QA benchmarks, two retrievers, and two frozen LLM backbones, this forward-only update improves a shared-interface baseline, transfers to held-out tasks, and approaches test-time gradient adaptation at much lower per-query cost

Remote sensing imagery supports diverse visual tasks, such as object detection and semantic segmentation, but these tasks vary across scenes, imaging modalities, and target objects. Since many remote sensing scenarios lack sufficient annotations, conventional supervised learning is often difficult to apply, motivating in-context learning as a flexible alternative.We present an in-context learning framework for remote sensing visual tasks, where a new task is specified by a single annotated example and applied to unseen images without retraining at inference time. Built upon diffusion models, our framework establishes target-reference associations through cross-attention and transfers the structural relationship between the reference image and its annotation to produce task-specific predictions. Unlike natural images, remote sensing objects are often small and provide weak semantic cues in early diffusion steps, making semantic alignment between target and reference images difficult. We address this limitation with a semantic-aware cross-attention that combines semantic and fine-grained details for more accurate cross-image matching. Remote sensing objects also appear at arbitrary orientations, causing large appearance variations and unstable in-context transfer. To improve robustness, we introduce a rotation-robust learning strategy that reduces sensitivity to orientation changes. Experiments show that our method achieves strong one-shot performance, establishing a flexible paradigm for adapting remote sensing models to diverse visual tasks.

Vision-language models (VLMs) learn strong statistical regularities during training, which can make them fail to perceive visual evidence when input images violate those regularities. Such failures are often treated as missing visual capability, but they may instead arise because the model has access to the relevant visual evidence and yet selects an answer dominated by learned training priors. In this work, we use controlled visual in-context learning to show that VLMs can overcome learned priors for visual tasks and use relevant visual evidence and capabilities. Specifically, we show that across five vision-centric tasks, demonstrations yield large gains on prior-conflicting counterfactual examples ($+27.2$% for Qwen3-VL 235B and $+18.1$% for Qwen3-VL 32B) while leaving real accuracy nearly unchanged. These findings also generalize across five VLM families with performance increases from $+9.7$% to $+30.3$%. To explain this effect, we analyze visual representations and find that demonstrations redirect attention toward grounding cues while reducing reliance on cues that support the learned prior. Finally, we find that in-context learning is limited when the underlying visual capability is weak or absent. Together, these results show that current VLMs possess visual grounding capabilities that standard evaluations may overlook, underscoring the need for controlled evaluation design in diagnosing and improving model capability.

Black-box optimization in science and engineering often comes with side information: experts, simulators, pretrained predictors, or heuristics can suggest which candidates look promising. This information can accelerate search, but it can also be biased, input-dependent, or misleading. Feedback-aware BO methods typically handle one task at a time, limiting their ability to generalize over multiple sources of feedback. In-context optimizers address cross-task adaptation, but usually assume that optimization history is the only available signal at test time. We study feedback-informed in-context black-box optimization (FICBO), where a pretrained optimizer conditions on both the observed history and cheap auxiliary feedback for the current candidate set. We introduce a structured feedback prior that models how feedback sources vary in their access, relevance, and distortion relative to the true objective, and use it to pretrain a feedback-aware transformer. At test time, the model estimates source reliability in context by comparing observed objective values with auxiliary signals, improving query selection. On synthetic and real-world tasks, FICBO effectively exploits informative feedback while remaining robust to weak or misleading sources, improving over other baselines. Empirical investigations further illustrate how the model perceives test-time sources, offering insights into its interpretability and decision-making process.

In in-context learning (ICL), a frozen pre-trained model solves tasks by conditioning on a prompt of a few input–output examples, without gradient updates. If the task was present in pretraining but the particular prompt sequence was not, the resulting in-distribution generalization is retrieval-based ICL. Learning-based ICL instead reflects out-of-distribution generalization: the model succeeds on prompts generated by a novel task. Empirically, both forms improve with scale. By analogy to benign overfitting in supervised learning, we call this in-context benign overfitting: larger models more faithfully memorize the pretraining tasks (improving retrieval ICL) while also generalizing better to novel tasks (improving learning ICL). We prove that this phenomenon already arises in a minimal in-context linear-regression feature-selection model. In contrast, standard in-context linear-regression models exhibit a retrieval–learning tradeoff, where the emergence of learning-based ICL coincides with degraded retrieval-based performance.

Running dense, open-weight large language models on commercially accessible workstations or single-GPU cloud setups is increasingly desirable due to cost and privacy constraints, particularly on tasks like coding and reasoning. Starting at around 24B parameters, however, they exceed the available GPU VRAM on these setups, forcing reliance on offloading methods that limit throughput, while leaving substantial CPU compute underutilized. Heterogeneous CPU–GPU speculative decoding is the dominant lossless-acceleration framework in this regime, but existing methods are limited by what we refer to as the **heterogeneity gap**: their CPU and GPU do not co-execute, their draft is a separate set of weights rather than a subnetwork of the verifier, and their VRAM footprint cannot scale to fill the available budget. We observe that channel-saliency methods induce an ordering on the transformer’s FFN channels so that the top-$k$ prefix closely approximates the full output, with graceful degradation as $k$ decreases. Building on this observation, we introduce **Speculative Tensor Parallelism (STeP)**, a training-free, self-speculative method that extracts a subnetwork of the verifier as a GPU-resident draft, places the remainder on the CPU, and verifies via concurrent CPU and GPU computation, provably preserving the verifier’s sampling distribution. Across dense models from 24B to 123B, GPU memory budgets from 32 to 192 GB, and five benchmarks, STeP outperforms SpecExec and SubSpec by up to $1.4 \times$ and $1.8 \times$ respectively, surpasses either tensor parallelism or speculative decoding alone, maximally utilizes the available VRAM, and closes the heterogeneity gap.

Non-convex optimization is central to machine learning and AI, yet unlike its convex counterpart, it still lacks a general and principled framework. In this work, we try to highlight the importance of combining noisy perturbations with over-parametrization mathematically under the classic setting of matrix completion (MC). Matrix completion is an important non-convex recovery problem in which only a subset of a matrix’s entries is observed, and the remaining entries must be recovered using low-rank constraints. This problem is notoriously difficult because observations are limited. However, we show that by injecting $\epsilon$-level perturbations into the nullspace of the observation mask, we can construct noisy matrix sensing surrogates whose solutions remain $\mathcal{O}(\epsilon)$-close to the ground truth with valid restricted isometry property (RIP) constants. The existence of RIP further enables powerful tensor frameworks to be applied with theoretical guarantees. This approach reflects a broader empirical trend in which stochasticity or noise improves the curvature of the optimization landscape, allowing over-parameterized (a.k.a large) models to better separate signal from noise. Assuming each matrix entry is observed independently, we establish quantitative recovery guarantees without explicit requirements on sampling rate or incoherence, and discuss how such assumptions can further strengthen our results.

Large language models (LLMs) are increasingly deployed in teams, yet existing coordination approaches often occupy two extremes. Highly structured methods rely on fixed roles, pipelines, or task decompositions assigned a priori. In contrast, fully unstructured teams enable adaptability and exploration but suffer from inefficiencies such as error propagation, inter-agent conflicts, and wasted resources (measured in time, tokens, or file operations). We introduce Language Agent Teams for Task Evolution (LATTE), a framework for coordinating LLM teams inspired by distributed systems, where processors must operate under partial observability and communication constraints. In LATTE, a team of agents collaboratively construct and maintain a shared, evolving coordination graph which encodes sub-task dependencies, individual agent assignment, and the current state of sub-task progress. This protocol maintains consistency while empowering agents to dynamically allocate work, adapt coordination, and discover new tasks. Across multiple collaborative tasks and a variety of base models, we demonstrate how LATTE reduces token usage, wall-clock time, communication, and coordination failures (e.g. file conflicts and redundant outputs) while matching or exceeding the accuracy of standard designs including MetaGPT, decentralized teams, top-down leader-worker hierarchies, and static decompositions.