NeurIPS 2026 Papers

Memory is becoming core infrastructure for long-horizon LLM agents, yet existing evaluations offer limited guidance on which memory substrate should be used under different operating regimes. We present a controlled harness evaluation of memory substrates for memory-augmented agents, covering dense and sparse indices, text records, structural stores, hierarchical stores, refinement-based memories, parametric updates, and activation-compatible context mechanisms. Across three backbone models and four benchmark suites spanning user-centric question answering and agent-centric decision-making, we instrument 26 performance and efficiency metrics under a unified harness. Our results show that no single substrate consistently dominates: broad retrieval benefits long-context factual QA, while excessive retrieval can harm sequential decision-making by shifting attention away from action-critical context. Scalability further introduces an additional routing axis, as substrates that perform well at moderate history lengths can become costly or brittle at longer horizons. These findings motivate substrate routing as a necessary component of adaptive agent memory systems and provide empirical guidance for designing efficient, reliable, and regime-aware long-term memory for LLM agents.

Long-term memory construction for LLM agents is hindered by two mismatches. The first is a **retrieval-objective mismatch**: existing systems usually retrieve memories for writing with a single holistic query derived from the current input. However, memory construction depends on several complementary aspects of the input, and a single query is insufficient to capture these editing needs. The second is an **optimization-granularity mismatch**: long-horizon memory learning is commonly optimized over full trajectories, whereas memory construction proceeds through incremental state transitions. Effective supervision should therefore assess how each local edit changes the memory state. We propose **GraphMemRL**, a graph-based memory construction framework that addresses both mismatches. GraphMemRL introduces edit-oriented query decomposition to retrieve complementary context for memory writing. It also uses action-native state-transition supervision to directly evaluate whether each graph edit improves the evolving memory state. To make long-horizon learning more tractable, GraphMemRL reformulates memory construction as factorized segment-wise optimization. The agent updates a persistent memory graph across segments under limited context budgets, preserving globally accumulated structure without full-history training. Experiments on long-memory benchmarks show that GraphMemRL improves question answering and produces better connected memory structures with more controlled memory growth. The gains are especially strong on multi-hop, temporal, and knowledge-update tasks.

Retrieval-augmented generation (RAG) divides into two camps failing in complementary ways on broad, long-tailed questions: iterative agentic RAG re-queries a biased distribution and drifts on prominent aspects, while structural RAG commits to a hierarchy fixed upfront and cannot admit aspects it initially missed. We propose GBT-RAG, which unifies the two by transposing gradient boosting into the text domain: each round is a query decomposition tree fitted to its predecessors' residual gap. This raises two challenges mirroring the pillars of boosting---natural-language critiques are not quantifiable, and successive rounds drift rather than fit the residual. We address the first with a text gradient, a structured residual in the retriever's embedding space whose support localizes missingness to specific leaves; we address the second by growing each tree only over under-covered leaves and admitting each rewrite through a monotone-improvement gate. Across standard benchmarks, GBT-RAG consistently outperforms the strongest iterative and structural baselines, with the largest gains on the long-tailed regime that motivated the design.

Enabling language model agents to autonomously adapt to new environments remains a fundamental challenge. While recent memory-augmented approaches have achieved promising results, they are largely heuristic in design and treat task execution and knowledge construction as decoupled processes. Inspired by how biological systems rapidly adapt through principled acting and learning, we propose \textbf{FEP-Agent}, a framework that grounds LLM agent self-evolution in the \textit{Free Energy Principle (FEP)}. Our system instantiates the two components of Active Inference into: a Worker agent that minimizes \textit{Expected Free Energy (EFE)} during online interaction, balancing goal-directed exploitation with curiosity-driven exploration to probe uncertain environmental dynamics; and a Builder agent that minimizes \textit{Variational Free Energy (VFE)} post-execution, updating a knowledge base while controlling complexity through counterfactual validation. To support scalable knowledge in open-ended environments, we introduce a structured semantic memory that employs associative linking and overlapping communities for efficient retrieval. The excellent results across multiple LLM-agent benchmarks demonstrate that FEP-Agent achieves principled, self-evolving systems.

Recent benchmarks for Large Language Model (LLM) agents mainly evaluate reasoning, planning, and execution. However, memory is also essential for agents, as it enables them to store, update, and retrieve information over time. This ability remains under-evaluated, largely because existing benchmarks do not provide a systematic way to assess memory mechanisms. In this paper, we study agent memory from a self-evolving perspective and introduce \texttt{EvoMemBench}, a unified benchmark organized along two axes: memory scope (in-episode vs.cross-episode) and memory content (knowledge-oriented vs.execution-oriented). We compare 15 representative memory methods with strong long-context baselines under a standardized protocol. Results show that current memory systems are still far from a general solution: long-context baselines remain highly competitive, memory helps most when the current context is insufficient or tasks are difficult, and no single memory form works consistently across all settings. Retrieval-based methods remain strong for knowledge-intensive settings, whereas procedural and long-term memory methods are more effective for execution-oriented tasks when their stored experience matches the task structure. We hope EvoMemBench facilitates future research on more effective memory systems for LLM-based agents. Our code is available at \url{https://anonymous.4open.science/r/EvoMemBench-2}.

Multimodal large language model (MLLM) agents increasingly operate over long histories of dialogues, images, and evolving facts, yet finite context windows and weak persistent memory cause forgetting, temporal inconsistency, and loss of precise past evidence. Existing multimodal memory agents either discard visual evidence by reducing images to captions or treat memory as a static store that is built once and never revised during reasoning; self-evolving agents further accumulate experience via prompting without tight coupling to retrieval and memory revision. Closing this loop with reinforcement learning additionally suffers from credit assignment difficulties over long, multimodal action trajectories. To address these challenges, we propose EvoMM, a reinforced self-evolving multimodal agentic memory framework that unifies memory construction, retrieval, and evolution within a single agentic loop. EvoMM maintains a mutable two-level memory of turn-level short memories and session-level long summaries, organized by a heterogeneous multimodal memory graph with temporal, semantic, image-co-reference, and keyword edges. An iterative PLAN→RETRIEVE→REFLECT→ANSWER policy then revises long memory on the fly and distills structured retrieval experience into a cross-query store for reuse. To optimize this policy, we further introduce PRR-GRPO (Planning-Retrieval-Reflection GRPO), which augments outcome rewards with gated, action-aware process rewards and performs dense credit assignment over a memory-action tree with dual-scale advantage normalization. Experiments on Mem-Gallery, MMLongBench, and LoCoMo show that EvoMM consistently outperforms strong text-only, multimodal, and agentic memory baselines, establishing a training paradigm for memory that reads, revises, and evolves with each query.

Large Language Models (LLMs) fail at a core aspect of human cognition: retrieving related episodes through shared elements **. When asked which clubs Ronaldo played for from 2008 to 2010, Claude Opus 4.6 answers Real Madrid, missing Manchester United, despite answering correctly when each period is queried separately. We study this failure through two complementary lenses drawn from cognitive neuroscience: *retrieval-based inference* (test-time recombination) and *integrative encoding* (encoding-time reactivation). "In vivo'', we benchmark frontier models (Claude, GPT, Gemini, DeepSeek, Grok, Mistral) and document 4--62\% binding failures on verified knowledge. Extended thinking, a proxy for retrieval-based inference, reduces but does not eliminate failures; multiple-choice probing recovers most of them, suggesting knowledge is encoded but not easily retrievable. "In vitro'', controlled experiments across model families and narrative lengths show that standard remedies offer partial progress: scaling improves binding but still falls short at 70B, extended training enhances binding but requires impractical redundancy, and transfer learning slightly helps with extensive supervision. Mechanistic analysis confirms correct answers appear in the model's top logits yet fail to surface during generation. Paralleling integrative encoding, we design Recall-based Synthetic Supervision (RSS), which prompts the model to recall past episodes and synthesize binding supervision online. RSS covers 92\% of evaluation queries, confirming the information is recoverable from parametric memory. Under realistic sequential training, however, hallucination, forgetting, and evolving supervision signals prevent this coverage from translating into accuracy. We establish episodic binding as an open problem in *parametric memory* ** Distinct from the perceptual binding problem; our usage draws on the associative inference literature on memory integration (Bunsey & Eichenbaum, 1996; Zeithamova & Preston, 2010).

Predicting transcriptional responses to genetic perturbations is fundamental to functional genomics and therapeutic discovery. Recent deep learning models have shown promise in single-cell perturbation response prediction, but they typically generate each response in isolation, without explicitly leveraging experimentally characterised responses to related perturbations. We introduce PT-RAG (Perturbation-aware Two-stage Retrieval-Augmented Generation), a plug-in retrieval-and-conditioning module for generative cellular perturbation response. PT-RAG augments an existing perturbation-response backbone with learned access to related perturbation contexts. The key challenge is that relevance is not fixed in this setting: functionally related genes may elicit different effects across cell types. PT-RAG addresses this with a two-stage retrieval mechanism: GenePT-based semantic retrieval first identifies $K$ candidate perturbations, after which a differentiable Gumbel-Softmax selector adaptively selects retrieved contexts conditioned on the control cell state, the query perturbation, and each candidate perturbation. Through cross-cell-type and cross-perturbation generalization tasks, PT-RAG consistently improves distributional similarity and often overall predictive quality; for example, on scGPT cross-cell-type results, Wasserstein distance drops by 10.7%. The code to reproduce our experiments is available at https://anonymous.4open.science/r/PT-RAG_NIPS-13D8/.

Retrieval-augmented generation evaluation checks whether model claims are factually grounded in retrieved documents. It does not check whether retrieved evidence is attributed to the correct entity. A clinical RAG response can pass every automated check (zero hallucinations, near-perfect faithfulness, real citations) while presenting drug $Y$'s clinical evidence as evidence about queried drug $X$. We term this **deceptive grounding** (DG): a failure invisible to faithfulness, hallucination, and citation checks because every claim is sourced from a real document, about the wrong entity. Using a controlled factorial benchmark across 13 models, we find DG rates spanning 8--87\% at peak adversarial conditions. Medical and biomedical fine-tuned models reach up to 86.7\%; domain specialization amplifies the failure rather than mitigating it. A controlled ablation identifies the mechanism: removing entity-specific clinical evidence from retrieved documents eliminates entity-attribution failure entirely, shifting all failures to confabulation. The two failure modes respond to the same trigger, taking different paths. Production measurement across 740 drug--disease pairs finds 7.8\% overall DG in a deployed RAG system, rising to 13.6\% for recently approved drugs. Entity-attribution verification (checking that cited evidence applies to the queried entity) detects DG at 92.3\% precision and 95.8\% DG recall; no existing framework implements it.

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.

Retrieval-Augmented Generation (RAG) improves the factual accuracy of large language models (LLMs) by grounding responses in external evidence. However, when retrieved context conflicts with models' internal parametric knowledge, LLMs may still generate answers that contradict the provided evidence, posing a key challenge to contextual faithfulness and the reliability of RAG systems. Motivated by recent mechanistic findings on the distinct propagation and progressive accumulation of parametric and contextual signals in LLMs, we propose Conflict-Suppressed RAG (CSRAG), a simple, training-free, decoding-time framework for resolving knowledge conflicts. CSRAG biases generation toward retrieved evidence by suppressing tokens associated with parametric knowledge while boosting tokens from the context via two complementary logits processors. Experiments on six challenging faithfulness benchmarks demonstrate that CSRAG consistently achieves state-of-the-art or near state-of-the-art performance across multiple backbone LLMs, while remaining fully training-free and lightweight.

Vector quantization is a fundamental primitive for scalable machine learning systems, enabling memory-efficient storage, fast retrieval, and compressed inference. Recent rotation-based quantizers such as EDEN, RabitQ, and TurboQuant have introduced strong guarantees and empirical performance, but the surrounding comparisons have been difficult to interpret because they rely on different distortion criteria, probability regimes, and implementation assumptions. As our first contribution, we provide a unified theoretical comparison of these methods and show that their relative advantages are criterion-dependent rather than absolute: TurboQuant is favorable for MSE distortion, EDEN is effective for expected inner-product distortion, and RabitQ provides strong high-probability control. This comparison clarifies the design principles behind each advantage and shows that no existing method uniformly dominates across the relevant measures. As our second contribution, we introduce *Block-Sphere Quantization* (**BlockQuant**), a new rotation-based block quantization algorithm designed around the spherical geometry of randomly rotated vectors. Unlike coordinate-wise quantizers, **BlockQuant** quantizes blocks on the sphere, preserving the geometry of rotated embeddings more faithfully. We prove that this block-spherical design improves expected distortion performance, including both reconstruction MSE and expected inner-product distortion. Experiments on real embedding data support these theoretical improvements.

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.

Retrieval-augmented generation (RAG) improves the factuality of large language models (LLMs) and vision-language models (VLMs) by grounding generation in external knowledge. However, most existing RAG frameworks assume a centralized retrieval document pool, which is often impractical in sensitive domains such as healthcare, where data are inherently distributed and subject to strict privacy constraints. Recent efforts on decentralized RAG primarily follow prompt-based paradigms that exchange raw, human-readable retrieved content, leading to substantial communication overhead and potential privacy risks. To address these limitations, we propose Representation-based Federated RAG (FedRepRAG), a decentralized RAG framework that avoids raw-data exchange during retrieval. FedRepRAG retrieves information from private local document stores and exchanges only compressed latent representations. To integrate the retrieved knowledge, we introduce a lightweight collaboratively trained projector that maps retrieval embeddings into the feature space of a frozen LLM/VLM backbone. This design makes the exchanged information interpretable only within the federated system, thereby enhancing privacy while substantially reducing communication cost. Experiments on decentralized visual question answering (VQA) and question answering (QA) benchmarks show that FedRepRAG consistently outperforms direct inference and local retrieval baselines, while substantially reducing retrieval tokens, computational cost, and inference-time communication overhead compared with raw-context transfer. Overall, FedRepRAG establishes an effective, efficient, and privacy-preserving framework for federated RAG.

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.

Backpropagation imposes fundamental constraints on training neural networks. It requires the retention of all forward activations until the corresponding backward pass completes, thereby preventing pipelined execution across layers. This work introduces Backpropagation-Free Local Antithetic Node Perturbation (BLANP), a fully local training framework that eliminates gradient propagation across layer boundaries while requiring no negative samples. Each layer updates its parameters from locally available signals through a zeroth-order antithetic perturbation estimator. To stabilize learning, the estimator is evaluated against a frozen exponential moving average critic, decoupling backbone updates from the concurrently evolving local classifiers. The antithetic perturbation formulation cancels even-order bias terms in the gradient estimate, yielding a first-order accurate update from two forward passes alone. A companion variant, BLANP-Exact, replaces the perturbation estimator with exact yet local gradients under a strict stop-gradient constraint, preserving a memory footprint that does not grow with network depth while recovering gradient fidelity. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 across MLP and convolutional architectures show that BLANP matches backpropagation on simple benchmarks. On the more challenging CIFAR-10, BLANP-Exact surpasses the best prior local learning method by 2.40% with a simpler architecture, and trails an identical backpropagation baseline by only 3.90%. The code is available at https://github.com/anonymous/BLANP.

Robust optimization (RO) provides a principled framework for decision-making under uncertainty, but its practical use is often limited by the need to manually reformulate uncertain optimization models into tractable deterministic counterparts. Recent large language models (LLMs) have been shown promising for automating optimization formulation, yet RO reformulation remains challenging because it requires precise multi-step reasoning and mathematically consistent transformations. To facilitate systematic evaluation of LLM-based reformulation, for which no dedicated benchmark currently exists, we develop AutoRO-Bench, a benchmark featuring an automated data generation pipeline for the core RO reformulation task and a curated dataset for the RO application task. To address the reformulation challenge, we propose Automated Reformulation with Experience Memory (AutoREM), a tuning-free memory-augmented framework that autonomously builds a structured textual experience memory by reflecting on past failed trajectories through a tailored offline adaptation procedure. AutoREM requires neither domain-specific expert knowledge nor parameter updates, and the resulting memory readily transfers across different base LLMs. Experimental results show that AutoREM consistently improves the accuracy and efficiency of RO reformulation across in-distribution datasets, out-of-distribution datasets, and diverse base LLMs.

Language agents increasingly operate over streams of related tasks, yet existing memory systems struggle to convert accumulated experience into reusable knowledge. Retrieval-augmented and structured memory methods record per-session observations effectively, but often couple acquisition and consolidation into a single online process, leaving the agent without a global view across sessions to discover recurring patterns, abstract shared procedures, or prune redundant entries. Inspired by complementary learning systems theory, we propose Auto-Dreamer, a learned offline consolidator for language-agent memory. Auto-Dreamer decouples fast per-session memory acquisition from slow cross-session consolidation. Given a region of a typed memory bank, the consolidator performs bounded tool-use to search memory, inspect candidate entries, and trace them back to raw source trajectories, synthesizing provenance-grounded replacement memories that supersede the original region. We train Auto-Dreamer via GRPO, using end-to-end agent performance as the reward signal to learn how to consolidate memories acquired through fast online experience. Trained on ScienceWorld trajectories alone, Auto-Dreamer transfers without retraining to held-out ALFWorld and WebArena, improves task success over fixed, RL-trained, and prompted memory baselines in both continual-memory deployment and fixed-bank consolidation, and does so with an active memory bank an order of magnitude smaller than competitive baselines.

Predictive coding (PC) is a local-learning alternative to backpropagation (BP), training deep networks via a local energy-minimization dynamics rather than a global backward pass. We introduce Augmented-Lagrangian Predictive Coding (PC-ALM), which maintains PC's inference budget but aligns each weight update toward BP by accumulating per-layer constraint errors into a layer-local Lagrange multiplier. In linear PC networks, PC-ALM converges to an equilibrium with exact BP gradients distributed across the network via only layer-local updates. We analyze PC-ALM in nonlinear PC networks up to depth 128 and show that it matches BP performance across all width-depth regimes, notably in deep narrow networks where PC underperforms. PC-ALM introduces recurrent dynamics in each layer's activations. Compared to PC's heat flow on a scalar energy, PC-ALM dynamics are driven by dual ascent on the Augmented Lagrangian. We observe ballistic credit propagation across especially deep networks, with prediction errors evenly distributed across layers, compared to PC's diffusive credit propagation. Beyond the algorithm itself, the augmented Lagrangian framework offers a generalization of PC, and may yield insights on how distributed systems could compute and propagate BP-like credit signals through purely local dynamics.