每天知道
你的领域
出了什么

Long-context inference with Large Language Models (LLMs) is bottlenecked by the linearly growing memory of the key-value (KV) cache. Existing compression methods reduce the cache through token eviction or approximation, but degrade sharply at aggressive compression budgets. We propose PatchKV, a training-free framework that compensates KV cache compression methods by carrying part of the context in the model's weights. PatchKV pairs an off-the-shelf compressed KV cache with a context-specific weight patch, which is computed once at context-loading time and served for downstream queries for the context. The weight patch is derived in closed form via ridge regression, by aligning the block-wise activations of context-derived reference query tokens under the full cache and the compressed cache. Once merged into the model, the patch leaves the forward graph and per-query inference cost unchanged in the single-context, multi-query setting. Across long-context QA (SCBench with up to 170K tokens, SQuAD, NIAH) and math (GSM8K) benchmarks on three model architectures, PatchKV consistently improves cache compression methods, suggesting an alternative direction to compensate them at aggressive budgets.

As LLM capabilities advance, agents are tackling increasingly complex tasks over longer horizons. Their growing interaction histories make memory compaction essential for staying within context windows and reducing prefill cost. Existing methods summarize the history or compress its KV cache, often adding model computation to preserve information for future requests. A new user request can change which history matters, but reassessing that history with the model requires re-encoding it if the KV cache has expired. Past attention provides signals of historical importance and dependencies between messages, while relevance to the current task must be assessed using the new user request. We introduce ReCAP, a memory compaction method that stores attention-derived importance scores and dependency links in a lightweight, persistent context graph. For each new request, ReCAP combines stored importance with relevance cues from the request and follows dependency links to select messages and their supporting context, without additional model calls for selection. Compared with Codex's default summarization-based compaction, ReCAP reduces estimated latency for compaction and cold restoration by approximately 95% on both Qwen3-Coder and gpt-oss. It also roughly halves the historical context per call on SWE-Together at comparable task quality and improves accuracy on the code tasks of Lost-in-Conversation over full history by 19.8 and 41.2 points.

Autoregressive video diffusion supports streaming generation and interactive control, but its KV cache grows continuously with the generated history. Existing compression strategies either discard history using fixed windows or select tokens through local attention and similarity signals, which do not directly measure whether the current chunk contributes information beyond the retained context. We introduce DeCoPrune, a training-free method that treats cache compression as a denoising-consistency problem. We find empirically that denoising difficulty provides a useful proxy for a token's value in long-term retention: tokens with larger step-to-final discrepancies tend to carry visual evidence that is less predictable from the retained context. DeCoPrune measures each current-chunk token's denoising difficulty using the discrepancy between its intermediate clean prediction and final denoised value, retaining high-discrepancy tokens in the long-term cache while pruning those with low discrepancy. To evaluate information retention, we introduce CMBench, comprising 58 approximately one-minute generated or real-world context episodes and 116 Reappear or Revisit continuation tasks that require recalling specific previously observed objects or scenes. Experiments with LingBot World v2 show that DeCoPrune preserves near-FullKV long-range recall while pruning over 85% of historical KV tokens and accelerating continuation generation by over $4\times$, substantially outperforming the evaluated compression baselines at comparable budgets. These results indicate that denoising consistency can serve as a model-intrinsic signal for retaining long-range information while reducing autoregressive inference cost. Our project homepage is https://decoprune.github.io. The code is available at https://github.com/DeCoPrune/CMBench, and the benchmark at https://huggingface.co/datasets/Aoraku/CMBench.

Large language model (LLM) serving requires substantial memory to store model weights and KV caches. As models grow larger and contexts become longer, memory capacity and bandwidth increasingly become bottlenecks for serving performance. Agentic workloads compound this pressure through repeated interactions over growing contexts, making it increasingly important to retain KV state for reuse. High-bandwidth flash (HBF) offers a way to expand accelerator memory capacity for large language model (LLM) serving, but its access costs and limited write endurance complicate its use. We evaluate HBF for high-throughput agentic serving across system design and scheduling choices to understand when additional capacity improves serving performance and energy efficiency. We introduce an HBM-HBF-host hierarchical storage system and buffered cache-aware scheduling, and use trace-driven simulations to analyze their effects on performance, energy consumption, and HBF write lifetime. Across the evaluated workloads, the fastest HBF-augmented systems reduce completion time by 36.1-87.7% relative to HBM-only systems. Modeled energy savings reach 59.1%, with benefits depending on the workload and weight placement. Buffered cache-aware scheduling extends estimated HBF write lifetime from 1.21 to 14.82 years in the evaluated configuration. These results demonstrate the importance of coordinating data placement and scheduling to improve serving efficiency while sustaining a practical HBF write lifetime.

We present Cascadia, a system for serving large language models on fleets of commodity Intel AIPCs using their CPU, integrated-GPU, and NPU resources. Every node embeds ingress, scheduling, and execution; inference requests require no dedicated routing control plane. Nodes join a libp2p QUIC mesh using CA-issued ed25519 admission certificates, gossip signed capabilities, exchange live load over direct peer streams, and route OpenAI-compatible requests to eligible peers. An operator-run certificate authority handles admission and fleet management outside the inference path. Three serving modes share one interface: whole-model execution on one node, load-balanced replicas, and pipeline-sharded chains using the compilation and speculative decoding mechanism of our companion paper. Optional KV-cache mobility reuses compatible conversation prefixes after a routing move, with cold recomputation on a miss. Signed response receipts and hash-chained logs support provenance and audit. A three-node Phi-3.5-mini NPU testbed delivered 3.10x the response throughput of its one-node configuration under ten concurrent requests; a separate four-node deployment recorded 4.06x the throughput of direct single-node serving. Paired latency observations, runtime measurements, and internal functional checks characterize the tested configurations. We compare Cascadia with IBM, Nutanix, VMware, and HPE platforms on deployment footprint, hardware requirements, scheduling, scaling, licensing, and trust, using vendor documentation. The paper repository provides benchmark scripts, curated measurements, and a claim-to-evidence map.

LLM inference systems may vary batch composition, prompt chunking, prefill/decode execution, and KV-cache reuse, eviction, or recomputation. These optimizations should not affect system outputs. Production systems, including vLLM's batch-invariant mode and SGLang's deterministic mode, target this goal but lack a formal system-level specification. We formalize deterministic LLM inference: under a fixed model and deployment configuration, requests with the same prompt and initial sampler state produce bitwise-identical logits at corresponding output positions across executions. Our tests find that these production modes produce different logits under some execution variations. To address this limitation, we present Vosti, an inference engine designed and verified against this specification. Vosti chooses kernels independently of runtime engine state and ties cached KV values to their logical token prefixes. Its proof decomposes at the engine/GPU kernel boundary: a Verus inductive proof establishes that scheduling and the paged, prefix-sharing KV-cache preserve output logits, while a Triton analyzer proves bitwise-equal selected kernel outputs across batches, query lengths, and paged KV-cache layouts. Vosti produces bitwise-identical logits across every tested execution variation and achieves performance comparable to vLLM's batch-invariant mode on decode-heavy workloads, while providing a stronger, formally verified determinism guarantee.

Production LLM serving stacks combine an inference engine's local prefix cache with a shared KV-cache tier for fleet-wide reuse. The local cache distinguishes requests by adapter, weight configuration and sharing domain, but the shared tier may key entries only by token content and coarse model metadata. This boundary erases provenance and lets identical tokens under incompatible computational or sharing contexts collide. We call this composition gap provenance-blind reuse and present its first systematic study. A source audit of three vLLM connectors confirms the structural omission, while runtime experiments reproduce it across vLLM and two SGLang releases, 12 models from 7 families (0.5 B-32 B), and over 160 configurations. Cross-adapter collisions reduce accuracy from 0.94 to 0.64, incompatible KV representations reduce reasoning accuracy to zero, and salt omission enables 93% prompt identification from timing. We formalize the missing guarantee as the KV provenance contract: for a declared dimension registry, shared keys must be injective over computational and sharing provenance, with identities stable across workers. Any dimension with a stable identity can therefore be added without connector-specific key logic. A canonical descriptor binds per-request and per-worker provenance into lookup and store keys, while a differential checker detects dimensions that change KV state without changing the key. Implemented in vLLM and SGLang 0.5.20 across three cache paths, provenance binding eliminates unsafe reuse while preserving legitimate sharing. Hit-path latency changes remain within 0.34 ms and below run-to-run variation; retention grows with provenance diversity.

Sparse attention is widely used to accelerate long-context inference in modern large language models (LLMs), but its input-dependent execution behavior introduces previously unexplored privacy risks. We identify a new GPU micro-architectural side channel, termed Sparsity-Induced Memory Access (SIMA), which arises from secret-dependent key-value cache access patterns induced by sparse attention. Based on this observation, we present SparLeak, a phase-aware side-channel attack that extracts SIMA traces during LLM inference and enables two practical privacy extractions: query attribute inference from prefill-phase traces and autoregressive response reconstruction from decoding-phase traces. By reconstructing approximate token-level sparsity profiles from page-level observations and applying profiling-based learning, SparLeak accurately recovers sensitive information, including user-query attributes and private LLM response content. Extensive evaluation across three LLM architectures, three sparse attention mechanisms, and three privacy-sensitive datasets shows that SparLeak achieves average attack success rates of 90.9% for attribute inference and 87.3% for response reconstruction under real-world LLM serving settings, highlighting the significance to account for SIMA leakage when deploying sparse-attention-based LLM systems. We provide anonymized SIMA traces, trained attack models, evaluation scripts, and documentation as artifacts at https://anonymous.4open.science/r/Janus_artifacts/.

Long-context LLM agents accumulate interaction histories that strain KV-cache memory and attention computation. Although sparse attention reduces these costs, heterogeneous cache representations and workflows hinder integration with existing inference engines, while prior sparse-serving abstractions support only specific layouts or workflows. We present SparseEngine, a ground-up, sparse-first inference engine whose shared lifecycle contract lets each method control its KV representation and computation while coordinating state transitions with common serving infrastructure. SparseEngine supports 15 methods across four categories and enables cross-request state management through Chain Cache, which resumes KV-eviction methods from retained history, and controllable Prefix-Cache Pruning, which removes KV from selected history regions while preserving logical-prefix matching. While maintaining method quality, SparseEngine delivers over 10x higher throughput with KV eviction, over 2.5x faster decoding at matched concurrency than vLLM, and over 2x end-to-end speedup on agent benchmarks. The code is available at https://github.com/CURRENTF/SparseEngine.

Language models are typically pretrained from random initialization. Recent work challenges this convention, showing that a brief warm-up on abstract, algorithmically generated data can provide a better starting point for subsequent learning of natural language. In this paper, we show that in small language models, such a warm-up improves specific capabilities that are not reflected in language-modeling perplexity. Our warm-up uses an abstract stack-manipulation task that requires compositional and state-tracking capabilities. Allocating as little as 1% of pretraining tokens to this data improves multi-hop question answering by up to 3.9 F1 points on MUSIQUE, with additional gains on HOTPOTQA and 2WIKIMULTIHOPQA despite comparable language-modeling perplexity. Controlled experiments show that the warm-up substantially accelerates the acquisition of deeper reasoning chains. We also explore what drives this transfer. First, the structure of the data matters: replacing the stack task with a queue fails to produce the same gains. Second, the gains are specific: performance improves on sequential reasoning chains, with no consistent benefit on tasks that combine or compare independent facts. Third, timing matters: mixing abstract data with natural language is far less effective than an initial dedicated phase, and exposure after pretraining completely removes the benefits. The early advantage persists through billions of subsequent language tokens. These results show that early abstract training can reliably shape the capabilities language models later acquire.

Rule-based reasoning, as in eligibility checks and contract reviews, requires language models to assess evidence against individual conditions and combine their judgments under explicit rules. Errors in evidence assessment can leave a decision unchanged, but misinterpreting or overlooking decision-critical evidence can reverse it. Identifying such evidence allows more capable models to focus on checking the corresponding condition judgments, supporting accurate and safe decisions. Recognizing the evidence's criticality requires understanding how evidence affects a condition judgment and how that judgment affects the decision. To achieve the goal, we propose a INTERvention-based imPACT learning framework (InterPact), which enables counterfactual verification of evidence criticality in rule-governed decisions. Specifically, its evidence intervention constructor generates training pairs for a propagation verifier by editing case facts with a frozen language model while holding rules and non-target conditions fixed. Human-reviewed labels record the resulting condition and decision changes, while complete state-to-decision mappings supervise consequences beyond the observed edit. During training, the verifier weights learned conditional decision predictions by evidence-based condition probabilities through a fixed composition operation, propagating decision-change supervision into the base model. At inference, the trained base model directly judges criticality from the original case and target evidence, without human or stronger-model supervision. On single-case evidence criticality verification over adapted rule-governed decision cases, InterPact achieves 68.28% accuracy, outperforming all six baselines. These results support learned decision sensitivity as a basis for prioritizing evidence checks.

The scale and expense of pre-training language models make efficient hyperparameter tuning essential, yet a principled guidance is still missing. In this work, we analyze language model pre-training dynamics from a local landscape geometry perspective. Our study reveals two distinct phases. In Phase I, sharpness of the local landscape is initially high, leading to instability and loss plateaus under large learning rates (LRs). The landscape shifts from sharp to flatter regions early in training. This dynamic explains the necessity of LR warmup and further suggests that larger peak LRs require proportionally longer warmup periods. In Phase II, the local landscape is governed by the gradient noise scale. Our theory identifies a depth flatness trade-off: high noise from smaller batches widens the loss basin, whereas reduced noise from larger batches deepens it. This theory motivates a dynamic batch-size (BS) scheduler that begins with a small BS and increases it late in training. Together, we provide a unified view of loss landscape evolution, which translates into actionable tuning strategies for large-scale pre-training.

Differentially Private Stochastic Gradient Descent (DP-SGD) is a leading approach for privacy-preserving fine-tuning of large language models (LLMs). Many decoder-only LLMs employ weight tying between input and output embeddings, a design choice originally introduced for parameter efficiency and improved language modeling performance in the non-private setting. However, the impact of weight tying under differentially private training remains largely unexplored. In this work, we investigate the role of weight tying in the DP setting using GPT2 and DistilGPT2 as representative decoder-only architectures. Interestingly, we find that untied embeddings consistently outperform weight-tied models under DP-SGD, achieving gains of up to 4.74% points in accuracy on SST-2, QNLI, and QQP. Beyond improved utility, untying embeddings enables the use of memory-efficient ghost clipping for DP-SGD. By contrast, weight tying introduces shared-parameter interactions that complicate standard ghost norm computation and largely negate its computational advantages. As a result, untied models achieve over 60% lower memory usage while preserving the benefits of ghost clipping. Our results indicate that untied embeddings provide a more effective and scalable design for differentially private training of decoder-only LLMs and highlight the need to revisit standard LLM architectural choices in the privacy-preserving setting.

Evolutionary search with large language models (LLMs) can stall when progress requires external knowledge the model lacks. Supplying relevant documents helps, but simply adding web search tool can keep returning the same pages as solutions change. We introduce EvoDuet, a bi-level optimization method that co-evolves solutions and search queries with fixed model parameters. At each iteration, a retrieval gate lets the LLM assess its knowledge gap and choose to retrieve new documents, reuse stored ones, or proceed without them. An inner loop refines queries and ranks documents by the solution scores they are predicted to yield; an outer loop generates candidates in parallel from these documents and records the evaluated outcomes for later searches. Across 21 optimization tasks with one candidate per iteration, EvoDuet raises OpenEvolve's normalized discovery gain from 74.1% to 78.0% with GPT-5.6-Luna and from 61.3% to 82.3% with Gemini-3.8-Flash, whereas Qwen3.5-9B does not benefit. Our best runs surpass the previously reported best scores on eight tasks, including Swap Reduction on Q20 and Rosetta, and match them on three more. EvoDuet also improves with other scaffolds (e.g., Top-K, EvoX) on Sums/Diffs and Denoising, demonstrating its applicability across evolutionary search scaffolds.

Predicting molecular properties for compounds that differ structurally from labeled training molecules is important for drug discovery and materials design. Tabular foundation models (TFMs) offer a promising approach through in-context learning, but their performance under structural shifts and the value of molecular comparisons in this setting remain underexplored. We study structural generalization in molecular property prediction and introduce MolPAIR (Molecular Pair-Augmented In-context Refinement), a framework that combines molecule-level and molecular-pair contexts without task-specific parameter updates. A global tabular foundation model (TFM) first predicts a query's property from labeled molecular examples. A second frozen TFM predicts differences in prediction errors between the query and labeled reference molecules, using these comparisons to refine the initial prediction. Across 58 MoleculeACE and Polaris tasks, CheMeleon representations combined with TabPFN-3 already outperform each evaluated baseline on a majority of tasks. MOLPAIR further improves this predictor on 46 of 58 tasks, with gains across four molecular representations and three TFM backbones. These results show that explicit molecular comparisons can strengthen tabular in-context learning for structural generalization while keeping the molecular encoder and pretrained model weights fixed. The code and datasets are available at https://github.com/nums-ai/MolPAIR.

Large reasoning models (LRMs) improve performance on complex tasks through extended reasoning, yet the same process can degenerate into redundant verification and persistent generation loops. Such uncontrolled reasoning increases inference cost and creates risks of resource exhaustion and service degradation. However, existing mitigations largely truncate long outputs or react to surface repetition, and thus fail to distinguish normal thinking from uncontrolled reasoning or explain how benign reasoning degenerates into harmful behavior. In this paper, we operationalize LRM generation as four states and further introduce Reasoning-state Analysis via Dynamic Attention Responses (RADAR), which identifies the current reasoning state in real time and characterizes how effective reflection can develop into uncontrolled generation. Guided by RADAR's analysis, we further realign abnormal attention distributions toward patterns observed in normal requests and examine how this correction affects excessive reflection and persistent looping. Temporal analyses show that uncontrolled reasoning is characterized by attention distributions that deviate from normal generation, with abnormal trends becoming detectable before repetition begins. Correcting these deviations through Attention Realignment consistently reduces looping while largely preserving benign performance. Together, RADAR provide a mechanistic account of how reasoning becomes uncontrolled, offering actionable guidance for identifying critical failure stages and designing targeted runtime interventions.

Recent advances in large language models have made automatic game generation increasingly feasible, yet reliably improving generated games beyond a playable version remains challenging. Naive iterative refinement can easily overfit a small set of test cases, producing fragile games with unresolved bugs, missing behaviors, and poor generalization to broader player interactions. We introduce RSIGame, an autonomous agentic game development framework with recursive self-improvement. RSIGame organizes development into complementary local and global loops. Concretely, a local explore-diagnose-improve loop broadly explores the executable game, diagnoses and prioritizes discovered issues, and performs evidence-grounded revision, where an evolving checklist continually accumulates new testing and improvement guidance. A global loop tracks overall quality, preserves the best checkpoint, and detects saturation or regression over long-horizon development. Beyond test-time improvement, RSIGame further internalizes successful development experience into the generator through training. Across 140 GameCraft-Bench tasks, two game engines, and five generators, RSIGame consistently improves game quality under matched development budgets. Notably, experience internalization enables Qwen3.8-27B to reach 61.38 on Godot and 58.53 on Phaser, exceeding GPT-5.5 one-shot scores while reducing Qwen's generation tokens by 11 times.

Looped Transformers repeatedly apply the same set of Transformer layers, giving them a recurrent architecture for latent computation. Their strong performance on iterative reasoning and length-generalization tasks suggests an appealing explanation: recurrence may provide an inductive bias that lets the model reuse a learned algorithm across loops. However, weight sharing alone does not imply that every loop performs the same operation. This raises a basic question: is each loop actually repeating the same computation, and if not, what routes the shared parameters to different operations? We study this question using graph walks as a test case. In the model's native trajectories, decoded predictions can advance by different numbers of graph steps or remain at a reached target, showing that recurrent progress need not follow a fixed one-loop-one-step pattern. We then show that a frozen loop can be steered toward different transitions by modifying its entering hidden state: a learned linear layer $J$ selects the desired transition without changing the shared Transformer layers. To test how this steering works, we use activation patching and find that attention patterns can recover its effects and switch the selected transition. Across five matched pairs of graph models, changing intermediate supervision during backbone training changes which transitions $J$ can induce. This suggests that $J$ selects computations learned by the backbone rather than creating new algorithms. Together, these results show that the hidden state can control shared computation, with attention routing as a causal pathway.

Improving text-to-image models has traditionally relied on increasing model size or the number of denoising steps. In this work, we explore an alternative way to scale computation by repeatedly running shared Transformer blocks within each denoising step, effectively increasing computational depth while keeping the parameter count fixed. This looped computation enables iterative refinement of internal representations without explicit reasoning tokens. However, naive looping fails to consistently improve image quality. We trace this problem to weak supervision across intermediate loops and unregulated attention updates that progressively erode local information. To overcome these challenges, we propose Looped Diffusion Transformer (Looped-DiT), which combines deep supervision across intermediate loops with self-modulating attention to stabilize looped feature updates. Under matched-parameter and matched-compute settings, Looped-DiT consistently outperforms non-looped baselines. Notably, a 260M-parameter looped model can surpass a model 6.5x larger across multiple text-to-image benchmarks while requiring 4.9x lower inference compute. Beyond this performance gain, we find that looped computation can offer a more effective form of iterative computation for diffusion models, with increasing loop depth yielding larger gains than adding more denoising steps under a fixed inference budget. Furthermore, deeper loops can progressively correct mistakes made in earlier loops, exhibiting behaviors suggestive of latent reasoning. Together, these results show that looped computation offers a promising way to scale visual generation models.

Causal AI is a branch of Artificial Intelligence which helps understand and reason about cause and effect relationships, not just patterns or correlations. Causal discovery aims to infer elements of the underlying causal structure--often represented as a directed graph--from observational and, when available, interventional data. While causal discovery is the fundamental step for moving beyond mere associations toward genuine understanding, and thus the basic building block of causal AI, it becomes intrinsically difficult when causal relations must be inferred from single observations. In such situations, standard causal discovery methods cannot be used and one has to identify causal relations from limited amount of information. This is typically the case for, e.g., sequences of events produced by different alarms which need to be analyzed on the fly to detect abnormal phenomena, which are usually rare. We show in this study that it is possible to leverage the predictive power of Large Language Models (LLMs) to infer causal relations between only two sequences of events. This approach, which is validated on both synthetic and real data, provides better results than standard causal discovery algorithms on several time series data, even though these data were converted into smaller, single observed sequences.

Organizations increasingly route employee feedback to leaders through large language model (LLM) summaries, an unaudited layer that silences already-spoken voice. We introduce a Voice Retention / Representation Ratio metric for representational bias in summarization and apply it to a bilingual (English/German) corpus of 2,586 free-text responses from a global professional service company. First, employees supply criticism more reliably than praise (withholding praise is 82 times more common). Second, across 45 leader-summaries the pipeline filters by popularity, not sentiment: criticism survives, yet a concern voiced once is dropped 86% of the time, with short and German-only content lost on the same axis (theme retention 0.14 vs 0.74; German directional). Controlling for frequency, sentiment has no independent effect; the harm is prevalence-driven, which sentiment-only audits miss. A targeted prompt recovers only named themes. We contribute the metric, field evidence, and a disaggregated voice-retention card.

Understanding the fundamental mechanisms of learning is essential for designing systems with strong generalization. Recent studies have shown that increasing the number of training domains, or enlarging the distribution shift among them, improves generalization when each domain contains sufficiently many data samples. However, the conditions under which the data samples can be considered sufficient remain unexplored. In this work, we fill this gap by establishing criteria for per-domain sample requirements based on the presented learning bounds. These criteria not only reveal an inverse linear scaling law between the number of training domains and the number of samples required per domain, but also explain the fundamental rationale behind the assumption of data sufficiency, thereby providing theoretical guidance for assessing the adequacy of existing datasets and constructing datasets. This differs from classical learning theory, as the number of samples required is highly dependent on the number of training domains. Additionally, we prove the close relationship between in-domain learning and out-of-domain generalization through the presented generalization bounds, and lastly discuss some key arguments.

Speech-to-speech systems must solve tasks whose requirements emerge across conversational turns. We introduce SpeechConversationBench (SCB), a focused evaluation of spoken mathematical reasoning using 103 sharded GSM8K problems. The framework compares the original problem delivered in one turn (full), its concatenated information shards delivered together (concat), and incremental spoken disclosure across turns (sharded). We report final-answer accuracy for four commercial speech systems and LEGO, a proprietary speech pipeline developed internally by the SCBX Innovation Lab team with explicit conversational context management. Relative to concat, sharded accuracy decreases by 5.0-25.3 percentage points across the four commercial systems. LEGO achieves 77.5 percent accuracy in all three conditions, compared with 76.6 percent sharded accuracy for GPT-4o Realtime. The two single-turn baselines distinguish sensitivity to problem reformulation from the additional challenges introduced by incremental spoken interaction.

Knowledge-intensive language-model systems typically represent external knowledge as text chunks or static graphs, with limited support for concept evolution, point-in-time reasoning, and distinctions between validated and inferred knowledge. We introduce the Concept Lifecycle Model (CLM), which represents concepts as persistent, graph-grounded, temporally versioned entities with explicit provenance and epistemic status, and Concept-Grounded Attention (CGA), which injects concept-graph structure into transformer computation through graph-biased self-attention (Form A) and gated cross-attention over concept nodes (Form B). We evaluate the framework in controlled settings using disabled-mechanism baselines. On 200 MuSiQue and HotpotQA questions with retrieval fixed, concept-graph retrieval recovers explicit multi-hop paths but does not improve evidence recall. Form A appears to steer attention, with 2.76 times more attention on gold than distractor concepts, but the same ratio occurs when Form A is disabled; the learned bias is negligible and no answers change. An identity-preserving Form B improves F1 from 0.188 to 0.221, but control concepts yield 0.213, indicating that most of the gain reflects added capacity. On LongMemEval, explicit temporal representation improves answer accuracy by 13 to 25 points across all tested generators, up to 122B parameters, while simplified CLM version resolution performs similarly to dated serialization because concept identity is not established reliably. On a synthetic source-independence task, protocol-derived epistemic status reduces unsupported assertions from 28% to 0.1% in a fine-tuned small model and from 19-68% to 0-5% in 72-122B models. Overall, the results support making temporal validity and epistemic status explicit, while showing that graph-attention diagnostics are not informative without disabled-mechanism controls.

Power-law learning curves are often treated as fixed properties of a model and its data, although learning-rate and batch-size schedules can change the observed loss. We study this dependence in noisy online SGD with linear random features. Conditional on the representation, an exact Volterra equation separates two response components: a forcing term that propagates unresolved target error and a memory kernel that propagates stochastic-error injections. We prove that either component follows a power law if and only if its cumulative weighted spectral mass has the corresponding low-spectrum scaling; individual eigenvalues and target coefficients need not obey coordinatewise power laws. Under a joint schedule, intrinsic time $T_t=\sum_{s<t}η_s$ controls optimization progress, while $r_t=B_t/η_t$ controls noise injection. Their interaction yields sharp conditions under which a schedule preserves, changes, or destroys the clean power law, together with a memory ceiling on noise reduction. The power-law random-feature model realizes this mechanism in $3+3(+2)$ propagation regimes with phase-dependent compute rates. Controlled nanoGPT experiments show that (1) learning-rate and batch-size schedules with matched $B/η$ paths are nearly equivalent in intrinsic time, (2) a forcing-memory surrogate accurately predicts loss across schedules, and (3) its fitted exponents across real-world datasets identify the regime of LLMs in $3+3(+2)$ map.

Steering large language models typically relies on linear, context-independent interventions in activation space, an assumption that recent work has challenged and that can induce an information bottleneck when a fixed representation must encode many behavioral distinctions. We introduce MetaSteer, a method that learns nonlinear interventions with context-dependent effects and applies them to attention projection matrices, producing activation effects that vary with the input context by construction and requiring no linear concept-geometry assumption. Framed as preference-based optimization, MetaSteer is trained once on a pooled preference corpus and transferred zero-shot to unseen concepts and out-of-distribution contexts. We find that, despite using low-rank adapters, MetaSteer induces structured, context-dependent changes in hidden-state trajectories while partially preserving aspects of their local trajectory dynamics, including velocity and curvature. We evaluate MetaSteer on three controlled text-generation benchmarks and three agentic settings across multiple model families and scales. MetaSteer matches or outperforms strong task-specific steering baselines on most aggregate comparisons in the zero-shot regime. Across the evaluated settings, stronger text-generation steering is associated with stronger agentic steering performance. We further discuss geometric trajectory effects, capability retention, and safety considerations raised by transferable steering.

Video Multimodal Large Language Models (Video-MLLMs) support reasoning over video inputs, yet remain vulnerable to jailbreak attacks that elicit policy-violating responses. Existing video jailbreaks primarily manipulate how harmful queries are visually presented, thereby treating video merely as a carrier. Consequently, the surrounding video scenario remains unexplored as a contextual attack surface. In this paper, we show that the same harmful query can elicit different safety responses when placed in different video scenarios. To systematically exploit this vulnerability, we propose SceneJail, an adaptive black-box jailbreak framework with two coordinated components. Adaptive Scenario Construction dynamically searches for a surrounding scenario that is contextually compatible with the harmful query. Scenario-aware Prompt Search uses black-box response feedback to search for textual guidance tailored to the selected scenario. Extensive evaluations on the HADES and SafeBench datasets across eight Video-MLLMs, including two proprietary models, GPT-4.1 and Gemini3.5-Flash, demonstrate the effectiveness of SceneJail. SceneJail-F, which presents the complete query persistently, achieves average attack success rates (ASR) up to 91.5%, outperforming the strongest baselines by 29.1 percentage points. Furthermore, SceneJail-S, which distributes the query across successive frames, remains highly robust against current defenses, retaining a 72.3% ASR even under strict image filtering.

Streaming video understanding requires models to process unbounded visual streams while preserving rich visual semantics across vast temporal horizons, posing a fundamental challenge for memory modeling. Existing approaches primarily focus on increasing memory capacity, either by compressing historical information into fixed-size representations or by extending storage beyond GPU memory. However, these methods largely rely on global or coarse-grained representations, inevitably losing fine-grained visual information. In this work, we argue that streaming video memory should explicitly encode structured and semantically meaningful representations, particularly at the entity level. To this end, we propose MEMO, a novel framework that models streaming video through multi-level, entity-aware structured memory. MEMO performs multi-level perception to jointly capture global semantics, entity dynamics, and spatial structures, partitioning streaming video into semantically coherent chunks. Each chunk is organized into a structured memory, where lightweight global and entity-level representations serve as retrieval indices, while the corresponding high-resolution visual content is retained separately for on-demand access. At inference time, MEMO performs query-specific retrieval over the structured memory and selectively recalls relevant visual evidence for downstream reasoning. Notably, MEMO is training-free and plug-and-play with existing multimodal large language models. Extensive experiments on StreamingBench and OVO-Bench demonstrate that MEMO consistently improves multiple base models and achieves state-of-the-art performance.

Emotion recognition in conversation has been widely studied, but applying Large Language Models (LLMs) to continuous dimensional emotion evaluation in multimodal dialogue remains largely unexplored. We propose an LLM-based framework that performs discrete emotion recognition and Valence-Arousal-Dominance (VAD) dimensional evaluation on IEMOCAP, incorporating acoustic cues as natural language descriptions following the SpeechCueLLM approach. We evaluate six models spanning the LLaMA, GPT, and Qwen families under zero-shot prompting, few-shot prompting, and LoRA fine-tuning. LoRA fine-tuned LLaMA models substantially outperform prompt-engineered GPT models on both tasks despite GPT's larger scale, a gap we attribute to domain adaptation rather than model capacity. Our best model achieves a Valence CCC of 0.7822, a new state-of-the-art on IEMOCAP. Ablation studies confirm that textual audio descriptions meaningfully improve smaller models (+3.5 to 3.6 weighted F1) while contributing little for the largest model, suggesting audio cues are most valuable when linguistic capacity is limited. The performance asymmetry across VAD dimensions closely mirrors the annotator agreement hierarchy in IEMOCAP's own annotations.

Looped transformers and Mixture-of-Experts (MoE) offer complementary routes to efficient scaling: recurrence increases computational depth at fixed parameters, while MoE sparsity expands total capacity at fixed active compute. Yet existing scaling laws model recurrence or sparsity in isolation. In this work, we introduce Loop Scaling Laws, the first scaling law to jointly model recurrence and sparsity alongside model size and data. At its core is a bounded, sparsity-conditional recurrence mapping that characterizes the effective-parameter gain from looping and how sparsity raises this gain. The laws predict the held-out loss of looped models more accurately than prior alternatives, and recover the standard dense and MoE scaling laws as special cases. Beyond prediction, the fitted laws provide a principled foundation for designing looped MoE models under compute and memory constraints. Downstream evaluations further demonstrate the complementary benefits of the two axes: sparsity delivers ~3x active-parameter efficiency, recurrence yields ~2x total-parameter efficiency on reasoning, and joint scaling further advances the performance frontier. As a practical extension, we show these gains hold at trillion-token scale: at matched training compute, a looped MoE with law-derived recurrence matches a ~2x larger non-looped MoE on the reasoning benchmarks, while enabling test-time scaling through recurrence.