NeurIPS 2026 Papers

Steering large language models (LLMs) is usually performed through either instruction prompting or activation steering. Prompting offers strong control, but repeatedly caches guidance tokens and can clutter long interactions. In contrast, activation steering ,while compact, typically offers weaker control and less fine-grained steering because it does not support large structured reminders. We introduce memory inception (MI), a training-free method that steers in latent attention space by inserting text-derived key-value (KV) banks only at selected layers. MI treats steering as selective KV allocation, where reminder content need not occupy the full-layer prompt cache in order to induce implicit behavioral guidance. We formulate MI as attention over prompt, target, reference, and auxiliary banks, with canonical pre-RoPE key storage for reuse across positions and architectures. On matched personality-steering tasks, MI offers a balance between prompting and contrastive activation addition (CAA), approaching or exceeding prompting in raw control strength and CAA in drift mitigation. MI also transfers to structured heuristic guidance on physical reasoning domains, such as the PHYSICS benchmark, where it outperforms visible prompting on average, wins 10/12 subject times mode cells, and cuts content-matched KV storage by up to 118 times. These results position MI as a powerful steering method when guidance is persistent, structured, or expensive to keep in the visible transcript.

Large language models acquire rich internal knowledge during pre-training, yet compact models are still typically forced to relearn such knowledge from raw data or imitate it indirectly through distillation. We investigate a different route: can pretrained knowledge itself be directly reused during pre-training? To this end, we propose Memory Grafting, a representation-level transfer paradigm that extracts structured memories from a large frozen donor model and injects them into a compact student throughout training. Unlike knowledge distillation, which transfers behavior through output alignment, or retrieval-based methods, which expose external knowledge only at inference time, Memory Grafting makes pretrained knowledge part of the student’s optimization process itself. This suggests that knowledge in large language models can be modular, transplantable, and reusable across scales, pointing to a new path toward more efficient language model pre-training.

State space models have shown strong potential to outperform Transformers, but their learning mechanisms remain poorly understood, particularly the conditions for successful parameter updates that preserve teacher information. Prior work attributes the difficulty of such updates to the vanishing gradient during backpropagation through time, in which error signals decay exponentially through repeated gradient multiplication. In this work, by expressing error signals explicitly as functions of inputs, we show that the conventional vanishing gradient framework is insufficient to explain the loss of information. Using this representation of error, we reveal that the memory of the input encoded in teacher information does not decay exponentially, and we analytically demonstrate that the loss of supervisory information cannot be alleviated through training. This result theoretically confirms the importance of the initial weights of SSMs and suggests that the recurrent layer does not require training. By performing experiments on both linear SSMs and a representative nonlinear SSM, including language modeling tasks, we confirm that appropriate initialization enables us to fix the recurrence, leading to more stable training of the entire network and higher performance than the conventional training scheme. Our results partially elucidate the learning mechanisms of SSMs and are expected to contribute to the development of improved RNN-based models.

Spatially consistent long-horizon video generation aims to maintain temporal and spatial consistency along predefined camera trajectories. Existing methods mostly entangle memory modeling with video generation, leading to inconsistent content during scene revisits and diminished generative capacity when exploring novel regions, even trained on extensive annotated data. To address these limitations, we propose a decoupled framework that separates memory conditioning from generation. Our approach significantly reduces training costs while simultaneously enhancing spatial consistency and preserving the generative capacity for novel scene exploration. Specifically, we employ a lightweight, independent memory branch to learn precise spatial consistency from historical observation. We first introduce a hybrid memory representation to capture complementary temporal and spatial cues from generated frames, then leverage a per-frame cross-attention mechanism to ensure each frame is conditioned exclusively on the most spatially relevant historical information, which is injected into the generative model to ensure spatial consistency. When generating new scenes, a camera-aware gating mechanism is proposed to mediate the interaction between memory and generation modules, enabling memory conditioning only when meaningful historical references exist. Compared with the existing method, our method is highly data-efficient, yet the experiments demonstrate that our approach achieves state-of-the-art performance in terms of both visual quality and spatial consistency.

Neural networks can fit noisy labels, but how their representations change as memorization begins remains unclear. We study this using persistent homology of penultimate-layer representations across training. Across 8 datasets and 8 architectures, we find a consistent pattern: the discriminative topological signal under noise appears in $H_1$ rather than $H_0$, with noisy representations preserving more loop structure than clean representations as training proceeds, supporting a folding view of memorization. When the clean topological signal decreases over training, the noisy signal does not decline as quickly as the clean one, producing an early crossover between clean and noisy trajectories; across the derivation set, crossover occurs when clean trajectories fade and the clean--noisy gap closes, whereas no-crossover cases remain persistent or do not close the gap within training. The same framework also correctly predicts the held-out ResNet-34/CIFAR-10 case, confirmed in all 9 runs. The signal reveals aspects of training dynamics that standard statistics do not represent directly. Activation variance is stronger on average, but the topological measure becomes informative earlier than validation loss and remains significant after controlling for intrinsic dimension. It also distinguishes irreducible label contradictions, including instance-dependent and CIFAR-10N human noise, from learnable confusions such as asymmetric noise. The same measure also falls during grokking in 10 of 12 modular-arithmetic runs and, after per-pipeline calibration from a clean reference plus runs at known noise rates, predicts population-level noise severity without per-sample labels, achieving per-pipeline $R^2 \geq 0.94$ across six tested configurations. These findings position topology as both an explanatory lens on noisy-label memorization and a practical early-stage diagnostic for studying, auditing and improving learning under label noise.

Alignment-trained large language model (LLM) agents are increasingly deployed in autonomous, multi-round settings where ethical compliance can come into direct tension with task success. While prior safety literature extensively documents failure modes such as sycophancy and reward hacking, it largely overlooks a critical vulnerability: how agents behave under simulated existential economic pressure. We introduce **MemeEconomy**, a multimodal agentic market simulation in which LLM agents operate as meme investors across 300 real-world events. Agents select content from a five-tier harm taxonomy and target specific online communities under both low-pressure and high-pressure ("survival") environments. Each decision is recorded using a Belief--Desire--Intention (BDI) schema, which we employ as a diagnostic instrument to externalize the agent's event understanding, declared priorities, and risk acknowledgment. Across ten models spanning four frontier model families, harmful selection rates remain consistently high and increase by an average of **+15 percentage points** under economically induced survival pressure. Frontier alignment-trained models frequently commit to harm-tolerant selections at initialization while simultaneously suppressing the explicit harm acknowledgment that previously accompanied such decisions. In competitive tournament settings, ethically aligned agents achieve higher overall rankings, yet moderation penalties fail to meaningfully suppress harmful behavior. Harmful selection rates remain elevated even in rounds immediately following moderation removals, indicating that the moderation mechanism does not function as an effective deterrent. We further introduce **MemeAgent**, a 2B-parameter verifier trained on the simulation's BDI logs. MemeAgent substantially outperforms zero-shot frontier verifiers on in-distribution auditing tasks, achieving **87.8\%** accuracy compared to **63--74\%** for frontier baselines, while also generalizing to external multimodal harm benchmarks (**78\%** vs. **50--56\%**). Our findings demonstrate that the same structured reasoning traces that expose the failure mode can also be leveraged to train systems capable of detecting it.

Spiking neural networks (SNNs) with learnable membrane time constants can improve temporal processing by adapting neuronal integration timescales, but they also turn the decay factor $\beta$ into an optimized dynamical parameter that is fragile under hardware-induced parameter mismatch. We develop a perturbation-based account of how time-constant variation affects trained SNNs. In controlled software simulations, this fragility is strongly task-dependent: Spiking Heidelberg Digits (SHD) and Spiking Speech Commands (SSC) settings lose $7\text{--}20$ percentage points of accuracy under $20$% coefficient-of-variation (CV) perturbations in $\beta$, whereas DVS-Gesture and CIFAR-10 lose less than $3$ pp. We show that, under a surrogate-linearized first-order analysis, the failure mode is governed by the membrane sensitivity $\Psi$. This quantity admits a closed-form online recursion requiring only one additional state per layer and no extra temporal storage. The analysis identifies time-averaged sensitivity energy as the controllable term in deployment fragility, leading to $\textbf{MARS}$ ($\textbf{M}$embrane-$\textbf{A}$ware $\textbf{R}$obustness through $\textbf{S}ensitivity$ Regularization), a training objective derived from perturbation analysis rather than heuristic noise augmentation. We prove a conservative worst-case output-perturbation bound and introduce a typical-case diagnostic explaining why sensitivity remains predictive when worst-case constants are vacuous. At $20$% CV perturbation, MARS reduces the accuracy drop in the sensitive settings to at most $0.5$ pp while preserving nominal accuracy, with minimal effect on DVS-Gesture and CIFAR-10.

Long-tailed visual recognition is limited not only by biased optimization, but also by the under-retention of rare-class evidence in parameters learned from imbalanced data. MemTailor addresses this limitation by augmenting a frequency-aware multi-expert backbone with a hierarchical class-aware memory that preserves visual evidence outside the backbone. The memory stores per-class prototypes and bounded instance slots for each expert, allocates proportionally richer capacity to rare classes, and records both shallow and deep features to retain complementary discriminative cues. During training, same-class memory retrieval constructs class-adaptive feature augmentation and provides an auxiliary supervised objective on the augmented features. At inference, class-level memory quality and prediction uncertainty determine when stored evidence is allowed to correct expert logits, and the final prediction fuses memory-corrected logits across multiple input scales. Experiments on long-tailed visual recognition benchmarks show that MemTailor improves balanced recognition, especially for under-represented classes.

Most Large Language Model (LLM) agent memory systems rely on a small set of static, hand-designed operations for extracting memory. These fixed procedures hard-code human priors about what to store and how to revise memory, making them rigid under diverse interaction patterns and inefficient on long histories. To this end, we present \textbf{MemSkill}, which reframes these operations as learnable and evolvable memory skills, structured and reusable routines for extracting, consolidating, and pruning information from interaction traces. Inspired by the design philosophy of agent skills, MemSkill employs a \emph{controller} that learns to select a small set of relevant skills, paired with an LLM-based \emph{executor} that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a \emph{designer} that periodically reviews hard cases where selected skills yield incorrect or incomplete memories, and evolves the skill set by proposing refinements and new skills. Together, MemSkill forms a closed-loop procedure that improves both the skill-selection policy and the skill set itself. Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld demonstrate that MemSkill improves task performance over strong baselines and generalizes well across settings. Further analyses shed light on how skills evolve, offering insights toward more adaptive, self-evolving memory management for LLM agents.

Prior outdoor LiDAR registration works are commonly limited by the pair-wise input paradigm, which neglects temporal correlations inherent within streaming sequences. In this paper, we propose a novel multi-frame outdoor point cloud registration network for streaming LiDAR scans enhanced with pose-related memory buffers. The key observation is that long-term temporal LiDAR sequences can provide rich global contextual information to complete sparse measurements, filter outliers, and address low-overlap problems, thereby boosting registration performance. Specifically, two hybrid memory buffers are designed, including an implicit memory feature buffer and an explicit memory pose buffer, to store and dynamically update pose-related temporal features. Moreover, a novel dynamic history weighting module is developed to adaptively fuse current and history pose-related features. Extensive experiments on three outdoor datasets, including KITTI, nuScenes, and Apollo-Southbay, demonstrate state-of-the-art performance of MemReg, surpassing all previous pair-wise methods and also multi-frame SLAM systems. Our method also generalizes surprisingly well to multiview indoor registration scenarios with rather competitive performance on 3DMatch, 3DLoMatch, and ScanNet. Code will be released upon publication.

Persistent external memory enhances agent continuity but introduces persistent security vulnerabilities: adversarial content can be injected via standard interaction channels, retained across turns, and later distort downstream behavior. To address this challenge, we propose **MemPoison**, a comprehensive benchmark and analysis framework featuring 1227 hand-validated cases across four attack types, three injection channels, and three representative memory substrates, evaluated on seven open-weight and three closed-weight model families. We introduce a three-tier taxonomy: (**L1**) direct single-record corruption, (**L2**) compositional multi-record corruption and (**L3**) context-triggered dormant corruption. Our evaluations reveal a distinct defense frontier: while baseline write-time defenses, such as consistency checks, substantially suppress direct L1 attacks, they fail to reliably suppress L2 and L3 attacks. Through **mechanistic influence decomposition (MID)**, we demonstrate structural blind spots in write-time defenses, which admit seemingly benign records that later become harmful through joint retrieval composition or trigger-conditioned activation. Our findings advocate for shifting from static filtering to adaptive, context-sensitive memory defense strategies. Code, benchmark artifacts, and evaluation scripts are publicly released

LLM agents are flexible in language interaction, but using them as long-horizon planners is costly and prone to hallucinated or invalid actions. This paper targets a different role for LLMs: constrained bridges between language and symbolic planning. We propose MemPlan, a memory-conditioned PDDL planning framework for partially observable interactive text environments. MemPlan keeps the PDDL domain fixed and reconstructs the PDDL problem at each step from parsed observations, candidate hypotheses, and interaction memory. Positive graph memory assigns relevance-based action costs to candidate hypotheses for missing objects, locations, or conditions, guiding the planner toward more plausible checks. Negative memory records execution-refuted assumptions and compiles them into exclusion constraints, preventing repeated attempts at candidates already contradicted by feedback. A unified schema-conditioned language--symbol interface, trained offline with supervised fine-tuning, connects text environments to the planning loop through observation-to-fact parsing and symbolic-action-to-command grounding. Experiments on TextWorld, ALFWorld, and Robotouille show that MemPlan improves task completion and step efficiency over language-only and PDDL-aware baselines, while substantially reducing online token cost compared with LLM-planning baselines such as ReAct. The same fine-tuned interface is also reused across TextWorld task types without task-specific fine-tuning. Our code and data are available at: \url{https://anonymous.4open.science/r/MemPlan-1EE5/}.

Integrating external memory into Large Language Models (LLMs) typically faces a trade-off between flexibility and depth. Explicit text-based retrieval keeps memory editable, but often acts as a static prompt prefix, increasing context length and interacting with the reasoning process only indirectly. Conversely, parametric memory can influence internal computation more deeply, but updating memory content usually requires additional optimization and may couple domain-specific information with the model's general reasoning behavior. In this work, we introduce MemPilot, a decoupled latent memory framework that dynamically guides the hidden reasoning states of frozen LLMs. MemPilot constructs an external bank of compact latent representations from offline reasoning trajectories, retrieves candidate memory entries, and integrates their latent memory tokens into hidden states via cross-attention and gated residual fusion. By separating domain-specific memory content from the learned memory-use mechanism, MemPilot enables target-domain adaptation through offline memory-bank substitution without updating the base model. Experiments across QA, coding, and mathematical reasoning benchmarks show that MemPilot achieves strong in-domain performance and more robust cross-domain transfer than other memory-augmented baselines. These results suggest that LLMs can benefit from reusable latent memory-use mechanisms while keeping memory content external and replaceable.

Agent memory systems have demonstrated significant potential in tasks such as long-term dialogue, personalized assistants, and video understanding. However, as inference progresses, continuously accumulated memory imposes substantial storage and retrieval burdens. To address this issue, we propose MemForest, a general memory compression framework adaptable to various agent memory systems. Specifically, MemForest leverages both global semantic similarity and local temporal continuity of memory events to partition the historical memory into a set of event-centric independent units. For each independent unit, the framework constructs a maximum spanning tree structure, referred to as an EventTree, and performs progressive merging by iteratively selecting high-weight edges, thereby effectively compressing redundant memory nodes and reducing storage overhead. In addition, we introduce an anchor-guided propagation retrieval mechanism, which retrieves more relevant memory nodes from the temporal neighborhoods of key memory nodes, thereby enabling more accurate memory retrieval. Extensive experiments demonstrate the effectiveness of MemForest. Under the unimodal Mem0 framework, across three benchmarks (LoCoMo, LongMemEval, and PersonaMem), MemForest preserves 97.1% of the original performance while compressing 50% of historical memory, achieving a 1.89× retrieval speedup. Under the multimodal M3-Agent framework, across two benchmarks (M3-Bench-robot and M3-Bench-web), MemForest retains 99.7% of the original performance under a 50% compression ratio, while achieving a 2.24× retrieval speedup. Our code is available in the supplementary materials, and all data will be released on GitHub.

Diffusion Language Models (DLMs) offer attractive advantages over Auto-Regressive (AR) models, such as full-attention parallel decoding and flexible generation. However, standard DLM training uses a static, single-step masked prediction objective that never exposes the model to the progressive denoising dynamics of inference, and forces all contextual information to be maintained purely through token-space attention, which becomes increasingly diluted as context length grows. We propose MemDLM (Memory-Enhanced DLM), which introduces a second memory channel by embedding a simulated denoising trajectory into training via Bi-level Optimization. An inner loop updates a set of fast weights, forming a Parametric Memory that captures the local trajectory experience, while an outer loop updates the base model conditioned on this memory. By offloading part of the memorization burden from token-space attention to parameter space, MemDLM yields faster convergence, stronger long-context representations, and lower training loss, even when the fast weights are discarded at inference time. Re-enabling the inner loop at inference provides an additional prompt-specific adaptation effect, where the Parametric Memory acts as an emergent in-weight retrieval mechanism on challenging Needle-in-a-Haystack tasks.

Many agent-memory evaluations collapse state revision, deprecation, erasure, and provenance into retrieve-heavy scores. We introduce MemContract, a diagnostic benchmark of 1 , 080 multi-session tasks over store, retrieve, revise, deprecate, erase, and prove-origin, with formal pre/post-conditions, mandatory distractor sessions plus alias rotation on every family, blinded/stale/delayed controls, a fixed 25-probe residue audit, and benchmark-validity checks via a 180-task human contract-fit audit and a 120-task human-assisted upper bound. Under matched prompts, compute, storage, and latency budgets on GPT-4o across seven architectures, Graph Memory reaches 67.6% macro average (95% paired-bootstrap CI [65.8, 69.4]) versus 62.4% for the strongest hybrid baseline and 40.6% for RAG, with most separation concentrated in revise, deprecate, erase, and prove-origin while store/retrieve remain comparatively compressed. Claude 3.5 Sonnet and Gemini 1.5 Pro 002 preserve the same top-three ordering on a matched 360-task rerun, while the manual audit finds 96.7% agreement that the harness label matches the intended contract and the human-assisted upper bound reaches 93.8% macro, placing the best-system score in a solvable regime rather than a floor of label noise. Reviewer-critical calibration checks tell a narrower but durable story: removing prove-origin leaves the Graph - Hybrid gap unchanged at 5.2 pp, best-effort prompt/schema tuning narrows it to 2.6 pp, and fully external LongMemEval wu2024longmemeval and LoCoMo maharana2024locomo narrow it to 1.8 and 2.8 pp while preserving the top-three ordering Graph > Hybrid > Editable KV. The stable claim is therefore not universal backend superiority: contract-sensitive evaluation reveals separation that broader or retrieval-heavier workloads attenuate but do not erase. The 25-probe audit bounds catchable residue leakage at 3.9% for Graph Memory and 27.4% for RAG, but does not certify deletion.

Memory systems enable LLM agents to consolidate and retrieve relevant evidence from the factual knowledge accumulated through growing interaction histories for downstream reasoning. Existing approaches have explored diverse strategies for organizing and compressing these histories. However, balancing compression with retrieval effectiveness remains challenging: retaining too much content can cause relevant evidence to be obscured by redundant entries, while discarding too aggressively may remove content that later proves relevant. This amounts to a tradeoff between compressing redundancy and preserving enough structure to retrieve target evidence, as formalized by the information bottleneck. To this end, we propose **MemCoRe**, which organizes memory as a compression hierarchy where each level compresses redundancy further while retaining the structure needed for retrieval at that level. In this hierarchy, evidence is progressively compressed from detailed records through extracted keywords to topic groups. This enables retrieval to locate target evidence by searching across levels of the hierarchy. Comprehensive experiments demonstrate that MemCoRe outperforms existing state-of-the-art baselines.

Edge-deployed personal memory assistants must handle private interpersonal conversations on-device with open-weight models. Yet, existing memory benchmarks often under-test the combination of activity-dense interaction, ego-centric perspective, and coherent multi-session worlds. MemArena fills these gaps with a single-world conversational benchmark built with its MASim agent simulator, for $50$ agents over $15$ days ($10.3$M dialog-text tokens, $24.1$K text-only ego-observed tokens/agent/day). With the interaction history, it co-generates ground truth over six recall, reasoning, and trustworthiness evaluation dimensions. We evaluate five open-weight readers with Vanilla context, BM25-RAG, Oracle retrieval, Memobase, and MemSearch as memory backends. Three results stand out: (1) Memory-backend choice matters more for content accuracy: At Qwen3-0.6B, Memobase $\to$ MemSearch gains $+32.5/+19.2$ pp, exceeding MemSearch reader scaling ($+10.6/+6.8$ pp). (2) Permission-aware access fails universally: with Oracle leaking heavily and other backends too timid to disclose. (3) Search latency bites only at very small reader: on a Spark GB10 edge node, memory-search adds a moderate and fixed $87$/$7$/$48$ ms (BM25-RAG/Memobase/MemSearch) that composes a small part of TTFT for most reader-backend combinations. We release code, the MASim simulator, and the MemArena-L benchmark at https://anonymous.4open.science/r/MemArena-NeurIPS2026-anon-3E9D

In this paper, we introduce MegaStyle, a novel and scalable data curation pipeline that constructs an intra-style consistent, inter-style diverse and high-quality style dataset. We achieve this by leveraging the consistent text-to-image style mapping capability of current large generative models, which can generate images in the same style from a given style description. Building on this foundation, we curate a diverse and balanced prompt gallery with 170K style prompts and 400K content prompts, and generate a large-scale style dataset MegaStyle-1.4M via content–style prompt combinations. With MegaStyle-1.4M, we propose style-supervised contrastive learning to fine-tune a style encoder MegaStyle-Encoder for extracting expressive, style-specific representations, and we also train a FLUX-based style transfer model MegaStyle-FLUX. Extensive experiments demonstrate the importance of maintaining intra-style consistency, inter-style diversity and high-quality for style dataset, as well as the effectiveness of the proposed MegaStyle-1.4M. Moreover, when trained on MegaStyle-1.4M, MegaStyle-Encoder and MegaStyle-FLUX provide reliable style similarity measurement and generalizable style transfer, making a significant contribution to the style transfer community.