Near-Oracle KV Selection via Pre-hoc Sparsity for Long-Context Inference

A core bottleneck in large language model (LLM) inference is the cost of attending over the ever-growing key-value (KV) cache. Although near-oracle top-k KV selection can preserve the quality of dense attention while sharply reducing computation and bandwidth, existing sparse methods generally rely on posterior heuristics, i.e., selectors conditioned on observed attention or proxy scores. Such conditioning introduces posterior bias: it tends to distort true token importance and miss salient tokens, thereby impairing long-range reasoning. To tackle this problem, we propose Pre-hoc Sparsity (PrHS), which selects KV entries before attention scoring and provides explicit accuracy control. Let the attention mass of discarded entries be delta (the dropped mass). Through a marginal-to-mutual-information analysis, we derive an upper bound on the mutual-information loss that depends only on the dropped mass. This relation explains failure modes of posterior heuristics and enables verifiable guarantees by controlling the dropped mass in advance. Within PrHS, we instantiate three orthogonal pre-hoc selectors along the axes of time, depth, and layer. Extensive experiments on LLaMA and Mistral families validate PrHS. Across GSM8K and CoQA, PrHS reduces retrieval overhead by over 90%, achieving 3x higher retrieval sparsity than HShare at matched or better accuracy. It incurs under 1% average degradation on LongBench, lowers attention FLOPs by about 15% versus prior sparse baselines, and yields a 9.9x speedup in attention-operator latency and 2.8x higher throughput on NVIDIA A100-80GB GPUs than the dense baseline.

MedCoG: Maximizing LLM Inference Density in Medical Reasoning via Meta-Cognitive Regulation

Large Language Models (LLMs) have shown strong potential in complex medical reasoning yet face diminishing gains under inference scaling laws. While existing studies augment LLMs with various knowledge types, it remains unclear how effectively the additional costs translate into accuracy. In this paper, we explore how meta-cognition of LLMs, i.e., their self-awareness of their own knowledge states, can regulate the reasoning process. Specifically, we propose MedCoG, a Medical Meta-Cognition Agent with Knowledge Graph, where the meta-cognitive assessments of task complexity, familiarity, and knowledge density dynamically regulate utilization of procedural, episodic, and factual knowledge. The LLM-centric on-demand reasoning aims to mitigate scaling laws by (1) reducing costs via avoiding indiscriminate scaling, (2) improving accuracy via filtering out distractive knowledge. To validate this, we empirically characterize the scaling curve and introduce inference density to quantify inference efficiency, defined as the ratio of theoretically effective cost to actual cost. Experiments demonstrate the effectiveness and efficiency of MedCoG on five hard sets of medical benchmarks, yielding 5.5x inference density. Furthermore, the Oracle study highlights the significant potential of meta-cognitive regulation.

DLLM Agent: See Farther, Run Faster

Diffusion large language models (DLLMs) have emerged as an alternative to autoregressive (AR) decoding with appealing efficiency and modeling properties, yet their implications for agentic multi-step decision making remain underexplored. We ask a concrete question: when the generation paradigm is changed but the agent framework and supervision are held fixed, do diffusion backbones induce systematically different planning and tool-use behaviors, and do these differences translate into end-to-end efficiency gains? We study this in a controlled setting by instantiating DLLM and AR backbones within the same agent workflow (DeepDiver) and performing matched agent-oriented fine-tuning on the same trajectory data, yielding diffusion-backed DLLM Agents and directly comparable AR agents. Across benchmarks and case studies, we find that, at comparable accuracy, DLLM Agents are on average over 30% faster end to end than AR agents, with some cases exceeding 8x speedup. Conditioned on correct task completion, DLLM Agents also require fewer interaction rounds and tool invocations, consistent with higher planner hit rates that converge earlier to a correct action path with less backtracking. We further identify two practical considerations for deploying diffusion backbones in tool-using agents. First, naive DLLM policies are more prone to structured tool-call failures, necessitating stronger tool-call-specific training to emit valid schemas and arguments. Second, for multi-turn inputs interleaving context and action spans, diffusion-style span corruption requires aligned attention masking to avoid spurious context-action information flow; without such alignment, performance degrades. Finally, we analyze attention dynamics across workflow stages and observe paradigm-specific coordination patterns, suggesting stronger global planning signals in diffusion-backed agents.

LAB-Det: Language as a Domain-Invariant Bridge for Training-Free One-Shot Domain Generalization in Object Detection

Foundation object detectors such as GLIP and Grounding DINO excel on general-domain data but often degrade in specialized and data-scarce settings like underwater imagery or industrial defects. Typical cross-domain few-shot approaches rely on fine-tuning scarce target data, incurring cost and overfitting risks. We instead ask: Can a frozen detector adapt with only one exemplar per class without training? To answer this, we introduce training-free one-shot domain generalization for object detection, where detectors must adapt to specialized domains with only one annotated exemplar per class and no weight updates. To tackle this task, we propose LAB-Det, which exploits Language As a domain-invariant Bridge. Instead of adapting visual features, we project each exemplar into a descriptive text that conditions and guides a frozen detector. This linguistic conditioning replaces gradient-based adaptation, enabling robust generalization in data-scarce domains. We evaluate on UODD (underwater) and NEU-DET (industrial defects), two widely adopted benchmarks for data-scarce detection, where object boundaries are often ambiguous, and LAB-Det achieves up to 5.4 mAP improvement over state-of-the-art fine-tuned baselines without updating a single parameter. These results establish linguistic adaptation as an efficient and interpretable alternative to fine-tuning in specialized detection settings.

Disentangled Representation Learning via Flow Matching

Disentangled representation learning aims to capture the underlying explanatory factors of observed data, enabling a principled understanding of the data-generating process. Recent advances in generative modeling have introduced new paradigms for learning such representations. However, existing diffusion-based methods encourage factor independence via inductive biases, yet frequently lack strong semantic alignment. In this work, we propose a flow matching-based framework for disentangled representation learning, which casts disentanglement as learning factor-conditioned flows in a compact latent space. To enforce explicit semantic alignment, we introduce a non-overlap (orthogonality) regularizer that suppresses cross-factor interference and reduces information leakage between factors. Extensive experiments across multiple datasets demonstrate consistent improvements over representative baselines, yielding higher disentanglement scores as well as improved controllability and sample fidelity.

Time Is All It Takes: Spike-Retiming Attacks on Event-Driven Spiking Neural Networks

Spiking neural networks (SNNs) compute with discrete spikes and exploit temporal structure, yet most adversarial attacks change intensities or event counts instead of timing. We study a timing-only adversary that retimes existing spikes while preserving spike counts and amplitudes in event-driven SNNs, thus remaining rate-preserving. We formalize a capacity-1 spike-retiming threat model with a unified trio of budgets: per-spike jitter $\mathcal{B}_{\infty}$, total delay $\mathcal{B}_{1}$, and tamper count $\mathcal{B}_{0}$. Feasible adversarial examples must satisfy timeline consistency and non-overlap, which makes the search space discrete and constrained. To optimize such retimings at scale, we use projected-in-the-loop (PIL) optimization: shift-probability logits yield a differentiable soft retiming for backpropagation, and a strict projection in the forward pass produces a feasible discrete schedule that satisfies capacity-1, non-overlap, and the chosen budget at every step. The objective maximizes task loss on the projected input and adds a capacity regularizer together with budget-aware penalties, which stabilizes gradients and aligns optimization with evaluation. Across event-driven benchmarks (CIFAR10-DVS, DVS-Gesture, N-MNIST) and diverse SNN architectures, we evaluate under binary and integer event grids and a range of retiming budgets, and also test models trained with timing-aware adversarial training designed to counter timing-only attacks. For example, on DVS-Gesture the attack attains high success (over $90\%$) while touching fewer than $2\%$ of spikes under $\mathcal{B}_{0}$. Taken together, our results show that spike retiming is a practical and stealthy attack surface that current defenses struggle to counter, providing a clear reference for temporal robustness in event-driven SNNs. Code is available at https://github.com/yuyi-sd/Spike-Retiming-Attacks.

A Step Back: Prefix Importance Ratio Stabilizes Policy Optimization

Reinforcement learning (RL) post-training has increasingly demonstrated strong ability to elicit reasoning behaviors in large language models (LLMs). For training efficiency, rollouts are typically generated in an off-policy manner using an older sampling policy and then used to update the current target policy. To correct the resulting discrepancy between the sampling and target policies, most existing RL objectives rely on a token-level importance sampling ratio, primarily due to its computational simplicity and numerical stability. However, we observe that token-level correction often leads to unstable training dynamics when the degree of off-policyness is large. In this paper, we revisit LLM policy optimization under off-policy conditions and show that the theoretically rigorous correction term is the prefix importance ratio, and that relaxing it to a token-level approximation can induce instability in RL post-training. To stabilize LLM optimization under large off-policy drift, we propose a simple yet effective objective, Minimum Prefix Ratio (MinPRO). MinPRO replaces the unstable cumulative prefix ratio with a non-cumulative surrogate based on the minimum token-level ratio observed in the preceding prefix. Extensive experiments on both dense and mixture-of-experts LLMs, across multiple mathematical reasoning benchmarks, demonstrate that MinPRO substantially improves training stability and peak performance in off-policy regimes.

Why Self-Rewarding Works: Theoretical Guarantees for Iterative Alignment of Language Models

Self-Rewarding Language Models (SRLMs) achieve notable success in iteratively improving alignment without external feedback. Yet, despite their striking empirical progress, the core mechanisms driving their capabilities remain unelucidated, leaving a critical gap in theoretical understanding. This paper provides the first rigorous theoretical guarantees for SRLMs. We first establish a lower bound that characterizes the fundamental limits of a single update step, revealing a critical dependence on the quality of the initial model. We then derive finite-sample error bounds for the full iterative paradigm, showing that performance improves at a rate of $\widetilde{\mathcal{O}}\left(1/\sqrt{n}\right)$ with sample size $n$. Crucially, our analysis reveals that the dependence on the initial model decays exponentially with the number of iterations $T$. This provides a formal explanation for why self-rewarding succeeds: it robustly overcomes poor initialization by steering the dynamics toward internal stability and consistency. Finally, we instantiate our theoretical framework for the linear softmax model class, yielding tailored guarantees that connect our high-level insights to practical model architectures.

SPA-Cache: Singular Proxies for Adaptive Caching in Diffusion Language Models

While Diffusion Language Models (DLMs) offer a flexible, arbitrary-order alternative to the autoregressive paradigm, their non-causal nature precludes standard KV caching, forcing costly hidden state recomputation at every decoding step. Existing DLM caching approaches reduce this cost by selective hidden state updates; however, they are still limited by (i) costly token-wise update identification heuristics and (ii) rigid, uniform budget allocation that fails to account for heterogeneous hidden state dynamics. To address these challenges, we present SPA-Cache that jointly optimizes update identification and budget allocation in DLM cache. First, we derive a low-dimensional singular proxy that enables the identification of update-critical tokens in a low-dimensional subspace, substantially reducing the overhead of update identification. Second, we introduce an adaptive strategy that allocates fewer updates to stable layers without degrading generation quality. Together, these contributions significantly improve the efficiency of DLMs, yielding up to an $8\times$ throughput improvement over vanilla decoding and a $2$--$4\times$ speedup over existing caching baselines.

Partial Feedback Online Learning

We study partial-feedback online learning, where each instance admits a set of correct labels, but the learner only observes one correct label per round; any prediction within the correct set is counted as correct. This model captures settings such as language generation, where multiple responses may be valid but data provide only a single reference. We give a near-complete characterization of minimax regret for both deterministic and randomized learners in the set-realizable regime, i.e., in the regime where sublinear regret is generally attainable. For deterministic learners, we introduce the Partial-Feedback Littlestone dimension (PFLdim) and show it precisely governs learnability and minimax regret; technically, PFLdim cannot be defined via the standard version space, requiring a new collection version space viewpoint and an auxiliary dimension used only in the proof. We further develop the Partial-Feedback Measure Shattering dimension (PMSdim) to obtain tight bounds for randomized learners. We identify broad conditions ensuring inseparability between deterministic and randomized learnability (e.g., finite Helly number or nested-inclusion label structure), and extend the argument to set-valued online learning, resolving an open question of Raman et al. [2024b]. Finally, we show a sharp separation from weaker realistic and agnostic variants: outside set realizability, the problem can become information-theoretically intractable, with linear regret possible even for $|H|=2$. This highlights the need for fundamentally new, noise-sensitive complexity measures to meaningfully characterize learnability beyond set realizability.