overfeed.news
Anuncio

SRExpert

Deja de saltar entre 6 herramientas de Kubernetes.

Monitoreo, alertas, seguridad, despliegues Helm y operaciones con IA en una sola plataforma.

Probar 30 días gratis

Tema

Seguridad y regulación

2470documentos

198últimos 7 días

arXiv AI Papers

Beyond Corrected Memory: Execution Consistency in Multi-Agent Systems

Shared memory coordinates agents' actions, but correct records do not establish that those actions satisfy task requirements. Memory governance and failure diagnosis regulate or inspect recorded information; they do not by themselves establish whether it is sufficient to judge task duties. We define execution consistency through duties governing state use, information handoffs, and final-state agreement, with explicit evidence conditions for judging fulfillment. Our core claim is that identical retained records can correspond to compliant and violating executions under the same task rule. Controlled removal of evidence such as receipt, action dependence, or response validity leaves 82.4% of opposite-label pairs indistinguishable; restoration separates 97.9% of the merged pairs. Natural-log annotations identify the defined violations in actual executions. However, existing logs do not always explicitly represent the execution relationships needed for these judgments. To assess the definition's practical value, we use CAVERT, a framework for consistency diagnosis and recovery, to extract supported relationships from logs and apply these criteria. It consistently outperforms contract-prompted LLM and rule-based baselines in diagnosis across all 12 benchmark-executor settings. Under the same gate and executor limits, it also outperforms rule-guided recovery in all four evaluated environments. These findings identify execution evidence that agent-memory and execution interfaces should preserve for reliable judgment.

en
arXiv AI Papers

When Tools Lie: Reliability of Mathematical Agents Under Corrupted Tool Feedback

Mathematical problem solving often requires deterministic computational steps that agents delegate to tools and implicitly trust. Yet tools can fail silently, returning plausible but incorrect results. How well can agents detect and correct corrupted tool call outputs? We study this through a controlled corruption framework where a hidden interceptor replaces tool call results with plausible incorrect information on targeted problems. We evaluate agents across 31 problems under four verification designs including no verification (baseline), mandatory same-context reflection, optional fresh-context verification, and optional structural verification. Without verification, corruption causes dramatic accuracy loss, from 100% down to 72.4%. Mandatory reflection fully recovers this performance to 100%. Optional verification improves accuracy only when models actively invoke it. Our results show that checking frequency is strongly associated with robustness differences, while unequal invocation prevents a controlled comparison of verifier quality. A supporting recovery experiment shows that full problem restart succeeds in 100% of cases after explicit detection. These findings demonstrate that verifier availability and verification policy are separate components of mathematical-agent reliability. Mandatory policies enforce verification while optional policies depend on the model's own choice to invoke it.

en
arXiv AI Papers

Natural Language Questions as an Interface for Knowledge Graphs: QRAKEN Graph Distillation and Semantic Self-Healing

Natural-language access to RDF knowledge graphs is a core Semantic Web ambition. Large language models (LLMs) have advanced Text-to-SPARQL, yet on unfamiliar graphs they often generate valid queries that misrepresent the populated data model. QRAKEN is a training-free, ontology-agnostic neurosymbolic pipeline grounding generation in empirical graph evidence rather than schema expectations. An offline distiller produces TTQL, a compact description of populated multi-hop patterns, conditional frequencies and path-conditioned literal examples, plus a class-property co-occurrence matrix. Online, TTQL guides the LLM, while deterministic syntax, vocabulary and data-model checks provide diagnostics for iterative refinement. On CK25 (First International Text2SPARQL Challenge), under matched-condition recomputation on a QLever snapshot, QRAKEN achieves strict F1 of 0.643 0.026 with GPT-4.1 mini and 0.652 0.012 with GPT-5.4: relative gains of 30% and 32% over the strongest recomputed participant, outperforming systems using the same base model family. Ablations identify TTQL patterns as the dominant driver (+0.31 strict F1 over a shape-only baseline); the refinement loop provides a cheap safety net, rejecting triple patterns unsupported by the co-occurrence matrix. Compared with auto-derived SHACL, TTQL yields 64% higher strict F1, supporting the value of empirical patterns beyond schema exposure. With two local 35B 4-bit open-weight models at zero marginal cost, the same pipeline matches the strongest recomputed participant, and TTQL advantages over shape-only and SHACL baselines persist. Results on a single, relatively small benchmark provide an initial empirical signal; monolithic TTQL injection on very open cross-domain graphs remains the main limitation.

en
arXiv AI Papers

When Plans Change Answers: Formalizing Cost-Accuracy Optimization for Semantic Queries

In semantic query engines, predicates are evaluated by machine-learned models, and the choice of a query plan affects not only the cost of a query but also its result. Existing systems either apply a fixed threshold to each semantic operator or tune accuracy per operator, without accounting for how errors propagate through joins. We give a formal problem definition for cost-accuracy optimization of such queries. Our starting point is the calibrated confidence that decision models such as Jev attach to each decision. It yields an expected error for every decision; weighting these errors by each decision's contribution to the output (in the simplest case, its fan-out) gives the expected output quality of a plan without any labeled data, and the same computation in reverse turns an output-level accuracy target into a price on each base or intermediate tuple. Building on this, we define an oracle semantics for relational algebra with semantic operators, physical plans as pairs of a logical plan and a decision policy, declarative output-level targets, and a hierarchy of plan equivalence. We show that accuracy is plan-invariant under pointwise-deterministic policies, and that selection pushdown is not quality-sound when escalation bands are calibrated on the plan's own candidates. Expected quality can be computed in polynomial time under bag semantics; under set semantics it follows the dichotomy of tuple-independent probabilistic databases when every relation carries a semantic predicate. Choosing which tuples to drop is NP-hard, while the optimization problem decomposes into per-tuple decisions through two Lagrange multipliers. Simulations on a synthetic workload illustrate these effects; an evaluation on real engines is left for future work.

en
arXiv AI Papers

DirectSpeech2LLM: A Simple End-to-End Framework to Mitigate Prompt Overfitting in Speech-LLMs

Speech-LLMs often exhibit prompt overfitting, where models solely trained on automatic speech recognition (ASR) instruction fail to generalize to new instructions such as speech translation and continue to behave primarily as ASR system. We propose DirectSpeech2LLM, a simple end-to-end framework that preserves the instruction-following ability of the LLM on unseen tasks when conditioned on speech. It computes distance-based CTC loss over the frozen LLM embedding matrix and uses greedy CTC labels to derive geometrically and temporally aligned speech embeddings respectively as an input to the LLM. Trained solely on 960 hours of LibriSpeech ASR data, DirectSpeech2LLM outperforms the cascaded system on ASR (seen task) and generalizes zero-shot to speech translation and emotion recognition (two unseen tasks), closely matching the cascaded system upper bound on these two new instructions despite seeing neither during training. We also find that geometric alignment strength plays a smaller role than previously assumed, as our modified CTC loss is shown to provide sufficient implicit geometric grounding without requiring an explicit regression loss. Results are consistent across two LLM families and scale with both more training data and model capacity.

en
arXiv AI Papers

POLAR: Ontology-Guided Risk Prevention for Tool-Calling LLM Agents

LLM tool-use agents operate in dynamic environments where many actions carry operational risk. However, most safety mechanisms react only after errors manifest. Existing pre-emptive approaches either fine-tune the agent on chain-of-thought deliberation or compile natural-language guardrails into runtime checks, but they do so without exposing a structural, auditable verdict. We propose POLAR, a guardrail framework for small tool-calling agents that assesses reversibility through a structured two-layer ontology. POLAR assigns each action a graded reversibility score by deriving a candidate inverse sequence; calls failing a threshold are pruned before execution. Evaluated on τ^2-bench across six agent models, POLAR improves mean task reward by 0.11 to 0.18 points on airline for four of six agents, but only eight of eighteen model--domain cells improve overall; retail and stronger agents often regress. POLAR provides an auditable structural check and characterizes its task-utility trade-offs. Reward is not a direct measure of prevented harm.

en
arXiv AI Papers

Self-Retrospection Distillation: Turning Post-hoc Experiences into Prior Foresight

Reinforcement learning with verifiable rewards (RLVR) turns agent experience into learning signals primarily through scalar outcome rewards after interaction. For group-relative objectives, however, this signal vanishes when all rollouts receive the same reward, even though their trajectories may reveal useful information about what the task requires and how the agent fails. We ask a complementary question: can hindsight teach an agent what it could have anticipated before acting? We introduce prospective learning, which uses post-hoc experience to supervise foresight predictions from the pre-interaction view, and instantiate it with Self-Retrospection Distillation (SRD). Intuitively, a completed trajectory reveals knowledge that would have been useful and pitfalls that should be avoided; SRD distills this privileged hindsight into trajectory-blind foresight of the same policy. Foresight serves only as a training target and need not be explicitly generated at inference time. Across 10 tool-integrated reasoning and long-horizon agentic tasks, SRD complements RLVR and self-distillation baselines with gains of up to 24.2 pp. Its advantage is especially pronounced when reward contrast is scarce: when 37--98\% of rollout groups are reward-uniform across model scales, yet SRD can still exploit learning signal from sampled trajectories. In the 2B setting, where 98\% of groups are all-failure, the RLVR training ends up at 0.0\% success, while adding SRD reaches 60.6\% under the same rollout budget. Our results suggest that post-hoc agent experience is useful not only for evaluating or improving behavior, but also for shaping predictive representations before available interaction.

en
arXiv AI Papers

Optimization Encoders: Rethinking Second-Order Meta-Learning for Neural Fields

Conditional neural fields represent signals continuously, but their effectiveness depends on how the conditional latent representations are inferred from observed data. In meta-learning, this encoding occurs through gradient updates induced by the decoder, tying representation learning directly to decoder design. We formalize this connection by interpreting latent optimization as an optimization encoder, unifying the roles of second-order differentiation, latent parameterization, and task supervision. This concept enables second-order meta-learning for end-to-end training of the encoding procedure alongside the decoder, and clarifies which learning pathway first-order approximations discard. Guided by this view, we introduce Attentive Latent Fields (MetaLF), an equivariant transformer-based neural field that contextualizes a latent pointcloud through self-attention. These interactions shape both field predictions and the updates that construct their representation, allowing local observations to inform coherent non-local structure. Disentangling the inner encoding objective from outer task supervision unifies reconstruction, classification, and segmentation within an end-to-end meta-learning framework, using reconstruction-only latent adaptation at test time. Controlled experiments on polynomial fields link latent coordination to lower effective rank and stronger alignment with the underlying function space. Across image and 3D shape reconstruction, MetaLF improves fidelity within three to five gradient updates, while supporting semantic prediction across images, shapes, and volumes. Together, these findings position the optimization encoder perspective as a unified basis for designing neural fields around how representations are constructed, coordinated, and used.

en
arXiv AI Papers

Same Feedback, Different Answer: Measuring Run-to-Run Instability in Frontier-Model Customer Feedback Analysis

AI agents are increasingly being programmed to automate knowledge work over large collections of unstructured data. Such automation requires repeatability: when the underlying evidence is unchanged, the agent's categories, priorities, and counts should not shift materially between runs, even if each individual answer appears plausible. We introduce a repeat-run evaluation framework that aligns semantically equivalent categories and focuses on two operating metrics: theme churn, the normalized change in the returned category set, and volume disagreement, the change in counts for categories that persist. We evaluate three recurring customer-feedback tasks across eight frontier models, corpus sizes from 100 to 5,000 records, multiple prompts, and three execution designs: raw generation, taxonomy-free hierarchical decomposition, and a taxonomy-grounded agent (TGA) using persistent themes, subthemes, and record-level predictions. With Claude Opus 4.8 and the 1,000-record corpus fixed, TGA reduces theme churn by 86--88% relative to both raw generation and hierarchical decomposition, while matched-theme volumes have zero disagreement. The taxonomy-grounded agent is more stable than every raw model in the screen, remains more stable at each corpus size, and keeps this advantage when theme matching is made stricter or looser. Although evaluated on customer feedback, the framework targets repeated synthesis of unstructured corpora more broadly, including financial reports, legal documents, incident records, and scientific literature. Overall, these results show that taxonomy grounding produces more consistent and repeatable outputs for recurring knowledge work.

en
arXiv AI Papers

Learning in Dreams, Winning in Reality: A Continuous Dyna Loop for a Ten-Hero MOBA

World models are usually judged from the inside: by prediction loss, by the return a policy earns in imagination, or by how convincing their frames look. We judge one from the outside. We learn a structured, multi-agent world model of a complete ten-hero MOBA (206 units, every hero acting every tick, games of up to 6,000 ticks), train a policy only inside it with 1,400-tick free-running imagined episodes, and measure that policy in the real game against the opponent the game ships with. The real game never provides a gradient; it provides the policy's own games as training data for the world model, and an online evaluation that selects and anchors the policy. Run as a continuous asynchronous Dyna loop, the policy wins 70.2% of real games as radiant (421 of 600; 95% CI 66.4-73.7) on seeds never used for any decision, up from 0% for dream training alone and 33.7% before the loop. It wins none as dire, and neither does the shipped opponent when it plays itself. Four findings explain the result. Model exploitation is invisible from inside the dream: every unanchored run collapsed within a few updates while no in-dream metric tracked the collapse. A world model that is accurate on its training corpus is badly wrong on the policy's own games, and Dyna repairs it there, which is worth +9.2 points of real win rate with the policy recipe held fixed. Finally, the policy inherits its world model's fidelity profile mechanic by mechanic: the model represents the macro game but not crowd control, cast timing or lethality, and the policy wins by map-wide pressure with almost no coordinated fighting. We release the world model, the dream-PPO harness, a world-model debugger, the evaluation protocol, and every policy and log.

en
AWS Machine Learning Blog

Introducing GLM 5.3 on Amazon Bedrock

GLM 5.3 from Z.ai is now available on Amazon Bedrock: a 753B-parameter mixture-of-experts model built for coding and long-horizon agentic tasks. Learn how to invoke it with the OpenAI-compatible APIs, cut cost and latency with prompt caching, and run an authorized security test with the open-source Strix agent.

en
arXiv AI Papers

Base Models Can Reason By Taking a Cue From Training Data

In this paper, we study how training data creates associations between the tokens at the start of a base model's response and the reasoning behavior that follows. First, we demonstrate that fixing particular starting token cues makes a base model's performance competitive with that of its reinforcement learning (RL)-trained counterparts on math and coding. For instance, the cue "." raises Olmo-3-7B's MATH-500 pass@1 accuracy from 42% to 78%, while "Alright," raises Qwen3-14B's from 72% to 87%. Second, RL makes these cues more likely, while fixing them recovers much of its performance gain over the base model. Third, we trace the reasoning effects of token cues to the training data. We perform causal data interventions to turn an arbitrary word, such as "chicken", into an effective reasoning cue, or remove an existing cue's effect. A similar edit makes the prompt instruction "Think duck duck goose" as effective as "Think step by step" at eliciting reasoning. We also find that the hidden state representations induced by different cues correlate with different document types from the training set. Finally, we extend our study of token cues with a case study in language model safety, finding that different cues elicit distinct refusal and compliance behaviors that correspond to different types of training data.

en
arXiv AI Papers

BiasFlow: Geometric Monitoring and Backbone Regularization for Spurious Feature Reliance

Worst-group accuracy (WGA) evaluates a trained predictor but does not characterize how its frozen backbone behaves when a new head is learned. We introduce BiasFlow, a hook-based toolkit for monitoring class-attribute centroid alignment (IBMI), within-class centroid separation (W-IBMI), and feature-projection sensitivity. IBMI is confounded by class-attribute correlation and is not a measure of causal feature reliance. We pair these diagnostics with BiasFlow Regularization (BFR), a supervised, composable class-conditional centroid-alignment penalty. W-IBMI verifies the quantity BFR optimizes; it is scale dependent and does not independently establish attribute removal. Across the reported small-scale benchmarks, adding BFR improves or preserves mean WGA, with gains up to +26.0 pp on UrbanCars. The principal independent stress test freezes CelebA-Std backbones and trains fresh heads on biased data: BFR+GroupDRO improves WGA from 40.7% to 64.1%, while Male probe accuracy decreases from 92.5% to 72.2%. Attribute information remains recoverable, and cross-task results are mixed. A controlled synthetic-watermark ImageNet experiment additionally improves watermark-shift accuracy by +23.0 pp under matched training. These results support evaluating centroid geometry and resistance to biased head retraining alongside WGA, within the tested protocols.

en
arXiv AI Papers

Learning to Read the Contextual Tokens in Diffusion Transformers

Multimodal Diffusion Transformers (MM-DiTs) jointly process visual and textual representations throughout generation. These models repeatedly update the text tokens through multimodal attention, forming dynamic contextual tokens whose function is not well understood. In this work, we introduce a framework for reading this contextual space through natural-language interrogation. We train a lightweight bottleneck network that maps intermediate contextual tokens into the input space of a frozen Large Language Model (LLM), allowing the LLM to answer questions about the emerging image directly from these hidden representations. Our reader reveals that contextual tokens encode a rich, global representation of the emerging scene: generation-specific semantics, including attributes left underspecified by the prompt, are accessible surprisingly early in denoising, while increasingly fine-grained details become readable over time. Remarkably, this information remains decodable even when the MM-DiT receives an empty prompt, showing that contextual tokens accumulate substantial image-specific information from the evolving visual representation itself. We further find that generations with more readable contextual representations tend to receive higher human-preference scores. Building on these observations, we introduce Contextual Alignment, a training technique that explicitly reinforces the visual-semantic information encoded in the contextual tokens, improving generation quality and distributional coverage. Together, our results establish contextual tokens as both an interpretable view into the internal dynamics of MM-DiTs and an effective target for improving generative models.

en
arXiv AI Papers

Recursive Video In-Context Learning for Agentic Robot

LLM agents that orchestrate frozen vision-language-action (VLA) policies improve across episodes through text memory, which records what the agent did but not how the task is done. A demonstration video shows it, but fits poorly into an agent's context. The full video slows every turn, fixed keyframes lose the contact detail that decides whether a grasp holds, and what the agent needs shifts from the task's structure while planning to the frames around each contact. We introduce Recursive Video In-Context Learning (RV-ICL), a training-free method that turns a demonstration into a hierarchy the agent navigates rather than a prompt it receives. The hierarchy is built from the sub-events of the demonstration, such as grasps and releases. Its levels grow finer, from keyframes of the whole task to phases, moments and short clips, and are exposed through read-only tools. The agent reads the coarse levels before planning. During execution it re-enters the hierarchy whenever a step needs more detail and loads only the clip of its current sub-goal. One demonstration per task is enough. Built on RPent, RV-ICL raises success from 92.6% to 96.5% on LIBERO-PRO and from 86.7% to 95.8% on LIBERO-Plus.

en
arXiv AI Papers

MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents

Memory has become integral to the LLM agent ecosystem, supporting information retention and reuse across interactions. However, most existing agent memory systems construct memory in a query-agnostic manner, which can incur unnecessary preprocessing cost and discard details that later prove essential. Recent studies have begun shifting memory processing toward runtime adaptation, but typically specialize in particular operations or fixed processing schemes, leaving flexible control over performance, cost, and latency largely underexplored. To address this challenge, we present MemPilot, a flexible framework that orchestrates on-demand memory curation under different performance--cost--latency preferences. Specifically, we optimize a multi-step LLM policy via reinforcement learning to iteratively choose between retrieving from query-agnostic memory and delegating query-specific curation of raw multimodal history to heterogeneous LLMs and VLMs. The policy jointly controls evidence amount, curation instructions, model selection, and visual access, enabling fine-grained allocation of runtime computation. To optimize this policy under competing objectives, we adapt objective-wise advantage decoupling by separately estimating each objective's advantage before aggregation. Moreover, we introduce prefix-based marginal utility estimation for fine-grained credit assignment across multi-step rollouts. Experiments on five multimodal agent-memory benchmarks demonstrate favorable performance--cost--latency trade-offs across optimization preferences, with preference sweeps yielding broader frontiers than existing trade-off-aware baselines.

en
arXiv AI Papers

Paradee: Distilling Kokoro-82M into an 8M-Parameter Single-Voice Text-to-Speech Model

We distill Kokoro-82M, a widely used open text-to-speech model with 54 voices, into Paradee, an 8.07M-parameter model that speaks one of them. Paradee keeps Kokoro's architecture with much narrower layers, and each of its two halves is trained separately against the frozen teacher. It has 10x fewer parameters and needs 15x less compute. We first synthesize a corpus with the teacher and keep its durations, pitch, energy and phoneme features. We then train a small text side to predict these values, and a small decoder to turn the teacher's saved values into the teacher's audio, first with spectral losses and then adversarially. Finally, we connect the two halves and quantize the weights to int8. It needs no alignment learning and no joint training, and it runs on one laptop. Stored in int8, Paradee is 8.5 MB, runs 25x faster than real time on one CPU thread, and scores 4.41 on UTMOS against the teacher's 4.52. The student initially kept a slight buzz, which we trace to the phase of voiced speech between 2 and 8 kHz. A phase-locking filter applied after synthesis removes most of it, with no training and no extra parameters. Code, model files and audio samples are at https://github.com/sahilmahendrakar/paradee

en
arXiv AI Papers

TAPDreamer: Transferable Adversarial Patches for World Action Models

World models learn to predict how their environment will evolve, making them an important foundation for general-purpose robotic control. Yet world action models depend on camera inputs whose manipulation can corrupt the visual representations used across tasks and action policies. Existing attacks on these models optimize against the victim's actions or predicted futures and therefore require access to target-model outputs. In this paper, we propose an attack, TAPDreamer, against world action models that instead uses a public encoder alone to construct a fixed local perturbation that transfers across tasks and action architectures. TAPDreamer requires no target-policy queries. Our key insight is that interactions between patch-induced changes in attention weights and value vectors broadcast a nearly identical representation shift far beyond the patch footprint, and this shift remains stable across task observations. Guided by this insight, TAPDreamer uses six frames from one source task to maximize the global L1 distance between clean and patched encoder representations. In closed-loop evaluation, one frozen patch per benchmark, covering about 6.5% of the input, reduces FastWAM's success rate from 97.7% to 0.0% across 40 LIBERO tasks and from 90.8% to 0.0% across 50 RoboTwin tasks; matched random patches retain 81.5% and 79.2% success. The same patches reduce success to 2.1% and 0.8% on two DreamWAM configurations and to 10.0% on Motus. These results show that protecting downstream action generation alone is insufficient: defenses for world action models must also secure shared visual encoders against persistent local perturbations.

en
arXiv AI Papers

Sharpen Without Search: On-Policy Distillation of Sequence-Level Power Distribution

A language model can give a correct answer more probability than any single incorrect answer and still usually sample an incorrect one, because the incorrect answers together hold more probability. The power distribution raises each complete answer's probability to a power above one and renormalizes, shifting probability toward answers the model finds most likely (sharpening). Sampling from it improves reasoning without changing parameters, but needs many scored candidates per query. We show that a model can instead be trained to produce such answers in one generation. On-policy power distillation (OPPD) runs a sequential Monte Carlo sampler in which the model being trained generates candidates and a frozen teacher's power distribution weights them; the same probabilities weight each answer in a maximum-likelihood update. Training raises single-generation accuracy by up to 23.0 points on MATH500 and 27.3 on GSM8K over the untrained model at the same temperature, and one generation scores 2.4 and 3.5 points above published power sampling with 64 candidates, recovering 94 percent of the gain that 16 candidates give the untrained model. For context, against GRPO trained with verified rewards from the same checkpoint and budget, OPPD scores 3.8, 4.0 and 5.4 points higher on MATH500, GSM8K and AIME using no reference answers; the two are complementary, and OPPD applied after GRPO adds up to 9.3 points. Trained only on mathematics, OPPD raises HumanEval accuracy by up to 5.3 points. One loss coefficient moves the sharpening exponent the model absorbs between 1.19 and 2.02, against 1.14 for ordinary on-policy distillation, and it rises mostly on the model's own answers. Gains hold across model families and sizes, including a model already trained with verified rewards, where lowering the temperature gives nothing and OPPD adds 4.4 points on MATH500. Code: https://github.com/ArminAzizi98/OPPD.

en
arXiv AI Papers

MC-Sparse: Deconstructing and Closing the Dense-Sparse Attention Gap in Diffusion Transformers

Sparse attention is a primary approach to reducing the latency of diffusion transformers in long-sequence generation tasks, such as video and high-resolution 3D asset generation. However, existing methods can degrade generation quality and fidelity at high sparsity levels. Through controlled oracle comparisons, we trace this degradation to three sources: constraints imposed by token grouping, inaccurate interaction selection, and the attention contributions lost when tokens are discarded. Guided by this analysis, we propose Meta-Cached Sparse Attention (MC-Sparse), a training-free framework that selects individual key-value (KV) tokens while organizing similar queries into tile-aligned groups for efficient GPU execution. MC-Sparse caches metadata comprising query groups, KV indices selected using exact attention probabilities, and residuals between dense and sparse attention outputs, and reuses them across subsequent denoising steps. Across video and 3D generation models, MC-Sparse achieves higher fidelity to dense-attention outputs and larger denoising speedups than existing sparse-attention baselines, without visible quality degradation. Relative to dense attention, it delivers a 1.80denoising speedup on Minimax-H3-Base and a 2.32speedup on 3D asset generation, both with negligible quality loss.

en