Papers

Voice activity detection (VAD) sits in front of almost every streaming speech recognizer, yet the choice of detector is often made by feel rather than by measurement. We report a frame-level offline benchmark of three detectors on the test-clean split of LibriVAD-small: the neural TEN VAD, the neural Silero VAD and a simple RMS energy baseline. The evaluation covers 2,214 noisy utterances (about 7.1 hours) built from LibriSpeech, nine noise families and six SNR levels from -5 to 20 dB. At a 16 ms hop, TEN VAD reaches a ROC AUC of 0.917, an average precision of 0.968 and an F1 of 0.913. A sweep over 20, 30, 40 and 50 ms windows shows TEN to be almost insensitive to frame length (AUC 0.916 to 0.919), while RMS stays near 0.75 to 0.77. Silero matches TEN only near its native 32 ms input size and drops to 0.63 at 20 ms. We hypothesize that this is likely caused by our fixed-frame protocol rather than by the model itself. Zero padding, sub-chunk averaging and unreset model state are the likely factors, but we have not ablated them. In a per-condition breakdown of TEN, babble noise at low SNR stays close to chance. We also show that F1 at a fixed 0.5 threshold is a weak metric on this data, because a trivial all-speech detector already scores 0.854. All settings needed to reproduce the results are given in the paper.

Routing policies in Multi-Agent Systems (MAS) represent valuable intellectual property, yet their exposure through execution logs creates significant risks of unauthorized policy cloning. Existing agent watermarking methods protect intellectual property by modifying internal generation processes, such as perturbing token sampling distributions or injecting structural mutations into action trajectories, which may compromise task fidelity or suffer from limited transferability across heterogeneous platforms. To address these limitations, we propose Decision Uncertainty-Oriented Watermarking (DUOW), a lightweight framework that operates at the external decision layer without modifying model parameters or generated outputs. DUOW introduces a plug-and-play scoring mechanism that evaluates candidate agents from semantic and operational perspectives, identifies functionally redundant routing decisions, and selectively redirects choices among equivalent alternatives to embed ownership signals. By coupling watermark behaviors with utility-preserving routing decisions, DUOW enables cloned policies to retain ownership signals during policy imitation while increasing the difficulty of watermark removal. A confidence-driven adaptive threshold dynamically determines embedding opportunities, balancing watermark capacity and task fidelity under diverse routing contexts. Extensive evaluations on four multi-agent routing benchmarks demonstrate that DUOW preserves high task utility, achieves reliable ownership verification under up to 90% log erasure, and introduces negligible computational overhead, with an average per-step latency of 0.043 ms and total embedding overhead below 0.54 ms.

Current evaluations of multimodal unlearning typically focus on whether a model stops outputting target information. This surface-level criterion, however, masks two critical failure modes. First, knowledge suppression techniques often induce confabulation, leading the model to hallucinate facts about the unlearned subject. Second, refusal training merely alters observable behavior without erasing the underlying knowledge; such models might decline open queries but still identify correct answers under multiple-choice probing. To overcome these limitations, we propose Forget-Aware Self-Distillation (FASD). This method is designed to achieve honest ignorance, a state where the model genuinely loses the target knowledge and learns to acknowledge its absence. FASD separates these goals into two complementary channels. The behavior channel distills appropriate refusals from an ignorance-prompted copy of the model itself. Crucially, target tokens are masked from the teacher distribution to avoid accidentally reteaching the forgotten concepts. Simultaneously, the knowledge channel directly suppresses ground-truth answer spans independently of the model's generated responses. Experiments demonstrate that FASD successfully unlearns target knowledge, prevents information leakage in open-ended tasks, and preserves performance on retained data. The resulting model not only forgets the specified information but accurately recognizes the boundaries of its own knowledge. Code and data will be released upon publication.

Consider an unknown graph $G = (V,E)$ with vertex degrees $d_v$, accessed only by sampling uniformly random edges and querying the degrees of their endpoints. We study the query complexity of estimating the number of edges $m$ and the degree moments $M_s = \sum_{v \in V} (d_v)^s$ for fixed integers $s \ge 2$. In this model, [Beretta and Tětek, TALG 2024] estimate $m$ within relative error $\varepsilon$ using $O(\log n/\varepsilon^2+\sqrt n/\varepsilon)$ expected queries without knowing the number of vertices $n$, and [Bhattacharya and Pradhan, CIAC 2025] estimate $M_s$ using $O(n^{1-1/s}/\varepsilon^2)$ queries when $n$ is given. We show that the query complexity of both problems is governed by the matching number $\nu$ of $G$ (the maximum number of pairwise disjoint edges) rather than the number of vertices $n$, and that the dependence on $\varepsilon$ for degree moments can be polynomially improved. Our algorithms rest on a simple observation: the endpoints of a maximum matching cover every edge, so a random endpoint of a random edge falls on this set of $2\nu$ vertices with probability at least $1/2$. For every $\varepsilon\in(0,1]$, we give advice-free algorithms that estimate $m$ and $M_s$ within relative error $\varepsilon$ with probability at least $2/3$, without knowledge of $n$, $\nu$, or other graph parameters. Specifically, our algorithms use $O(\min\{\nu/\varepsilon, 1/\varepsilon^2+\sqrt\nu/\varepsilon\})$ expected queries for $m$ and $O(\min\{\nu^{1-1/s}/\varepsilon^{2/s}, \nu^{1-1/(2s-1)}/\varepsilon^{2/(2s-1)}\})$ for $M_s$. Finally, our lower bounds show that for every $k \ge 2$ and sufficiently small $\varepsilon$, these query complexities are optimal up to constant factors over graphs with matching number at most $k$.

Due to the high cost of hardware experiments, robot policies are predominantly evaluated by comparing their empirical performance on a limited number of hardware trials. However, finite-sample statistical uncertainty can lead to erroneous conclusions, especially when ranking multiple policies using a fixed batch of evaluation data. We propose AnyRank: a sequential method for multi-policy ranking that (i) adaptively allocates trials to hard-to-separate policies, (ii) provides (partial) rankings at any stopping time while bounding the probability of an incorrect ranking, and (iii) optionally exploits prior information to prioritize hardware evaluations more efficiently. This prior information could be informed from expert intuition, a small set of held-out hardware evaluations, or from policy rollouts in a simulator or an action-conditioned video world model. Notably, the prior can be imperfect; while a strong prior can provide efficiency gains on hardware, any inaccuracies in prior policy rankings do not invalidate the statistical assurances of AnyRank. We achieve the aforementioned properties through novel theoretical contributions that combine sequential testing with multiple-hypothesis graphical testing approaches. Experiments in simulation and real-world policy evaluation benchmarks show that AnyRank requires up to 50% fewer hardware evaluations compared to existing statistically rigorous approaches, while guarding against the incorrect rankings produced by heuristic methods used by state-of-the-art benchmarks.

In this work, we aim to develop an embodied MLLM agent that understands and solves embodied tasks by learning, abstracting, and reusing transferrable and generalizable behavioral patterns. To this end, we propose StateSkill, a new skill representation for embodied agents, which introduces the concept of "finite-state machines" into skill modeling for effective and efficient skill construction and learning. The core idea of StateSkill is to model embodied interaction experiences as a finite-state machine, where a skill is organized around a set of task-progress states, state transitions, and state-conditioned guidance, enabling embodied agents to dynamically adjust their skills as the task progresses while keeping skill contexts compact and stage-relevant. We further develop StateSkill-Evolve, an experience-driven evolution mechanism that iteratively refines StateSkill through retrospective analysis of new successful and failed state-guided trajectories. Upon this, we introduce StateGRPO, which integrates StateSkill evolution and model optimization via state-conditioned GRPO into a unified iterative learning loop, enabling state-conditioned skills and model capabilities to co-evolve and reinforce each other. Extensive experiments demonstrate the effectiveness of our proposed methods and trained model (i.e., StateSkill-Agent) across various embodied tasks.

We consider an elliptic Cauchy problem for the Poisson equation in which both Dirichlet and Neumann boundary data are prescribed only on a portion of the boundary, while no boundary conditions are available on the remaining part of the boundary. Such problems are severely ill-posed. In this work, we study the convergence of an iterative method based on the alternating imposition of Dirichlet and Neumann boundary conditions. We formulate the scheme within a Hilbert space framework, where a yet another alternation becomes apparent: each iteration can now be interpreted as an orthogonal projection onto an appropriate closed subspace. This perspective provides a transparent geometric interpretation of the algorithm and yields a concise convergence proof based on von Neumann's classical result on alternating (orthogonal) projections. In addition to elucidating the convergence of the method, this alternating projections perspective also offers a general rate of convergence in terms of the Friedrichs angle (between two closed subspaces). We point out by means of an example that such convergence rate estimate degenerates for the elliptic Cauchy problem under consideration.

The effective integration of features derived from RGB images and 3D point clouds presents a substantial challenge in the field of multimodal anomaly detection (MAD). Existing MAD methods typically rely on static fusion paradigms built upon frozen pre-trained encoders. While these frameworks facilitate spatial alignment, they lack the mechanism to establish latent semantic correspondence, thereby inducing cross-modal interference that degrades anomaly discriminability. To mitigate these limitations, this study proposes a Bidirectional Adaptive Transformers (BAT) framework, which constructs comprehensive and semantically aligned multimodal representations by leveraging complementary attributes across modalities. The BAT framework introduces a Bidirectional Feature Alignment (BFA) module, which employs a mutual predictive mechanism to bridge the gap between 2D textural patterns and 3D geometric structures. This bidirectional interaction explicitly enhances the textural granularity of point cloud descriptors while augmenting the spatial contextual awareness of RGB features. To ensure that the latent representations possess sufficient informational density for alignment, BAT incorporates a parameter-efficient adaptation architecture. By utilizing lightweight image and point cloud adapters, the framework facilitates specialized fine-tuning of pre-trained models, mitigating the information loss inherent in frozen encoders. Extensive experiments conducted on MVTec 3D-AD and Eyecandies confirm that BAT achieves state-of-the-art anomaly detection performance, significantly outperforming existing MAD methods.

Navigating cluttered environments while carrying objects remains a challenging problem for quadruped robots. Compared with works that focused on load-aware locomotion adaptation to maintain gait stability, existing autonomous navigation systems generally ignore the influence of carried objects' geometries and use fixed footprint inflation, leading to overly conservative navigation. To fill this knowledge gap, we propose a two-phase load-aware navigation learning framework. In Phase 1, we train LOADNav, a load-aware navigation policy via reinforcement learning, enabling the robot to adapt its trajectory and clearance behavior according to the geometric size of the carried object. In Phase 2, we train CoGE, an online geometry predictor via imitation learning, which estimates unknown payload geometry from occasional contact events. We evaluate the framework in cluttered environments with loads of varying sizes and shapes. The results show that our navigation policy exhibits clear load-aware behaviors, and the online geometry predictor estimates object geometry from occasional contact events, refining its estimate and adapting the navigation behavior accordingly. Without requiring any prior knowledge of the payload, it narrows the gap to geometry-aware policies.

Given $n$ nonnegative integer weights $w_1, \ldots, w_n$ and an integer knapsack capacity $T$, the $\#$Knapsack problem asks to count the number of subsets $I \subseteq \{1, 2, \dots, n\}$ such that $\sum_{i \in I} w_i \le T$. A long line of research has been devoted to designing faster $(1 \pm \varepsilon)$-approximation algorithms for this \#P-hard problem. Until very recently, the best-known time complexity was $\tilde{O}\bigl(n^{2.5}/\mathrm{poly}(\varepsilon)\bigr)$, achieved by the randomized algorithm of [Dyer, STOC 2003] and the deterministic algorithm of [Gawrychowski, Markin, and Weimann, ICALP 2018]. The recent work by [Feng and Jin, SODA 2025] improved the randomized complexity to $\tilde{O}\bigl(n^{1.5}/\mathrm{poly}(\varepsilon)\bigr)$, leaving open the question whether this $n^{2.5}$ barrier could also be broken by a deterministic algorithm. In this work, we answer this question in the affirmative by presenting a deterministic $(1 + \varepsilon)$-approximation algorithm for \#Knapsack in $\tilde{O}\bigl(n^{2.4}/\mathrm{poly}(\varepsilon)\bigr)$ time. Our key technical innovation is to deterministically simulate the balancing technique, which was previously used in the literature of knapsack-type problems via random partitions and seemed to require randomization.

Spiking Neural Networks offer event-driven, energy-efficient computation for speech command recognition, yet most existing methods follow fixed computation paths, allocating redundant computation to easy samples while providing limited insight into how evidence evolves during inference. To address these challenges, we introduce SpikCTU, a compact spiking network with adaptive recurrent computation for speech command recognition. Specifically, a hybrid spiking embedding module preserves continuous input information alongside spike features to construct a fixed temporal memory. A shared Continuous Thinking Unit (CTU) repeatedly reads this memory using keys derived from spikes and continuous values, and updates a binary recurrent state with lightweight stage adapters. Furthermore, combining a fixed base prediction with accumulated corrections at each step enables learned aggregation over all steps or confidence-based early exit using only the predictions available so far. Experiments on three speech command recognition benchmarks show that SpikCTU achieves competitive accuracy with only 0.16–0.17M parameters, and the early exit strategy can substantially further reduce computational cost without compromising accuracy. In addition, trajectory and attention analyses reveal that recurrent steps enable error correction and progressively refine predictions. Code is available at https://github.com/Rivflyyy/spikctu.

Wheel odometry provides an efficient motion prior for ground robots but is prone to transient violations such as wheel slip or terrain-induced tilt. In filtering-based odometry, these violations frequently trigger divergence when they coincide with constrained geometric observability (e.g., featureless corridors). We address this challenge by proposing ToF-Wheel Odometry (TWO), a tightly-coupled filtering framework that integrates wheel kinematics with geometric constraints from Time-of-Flight (ToF) cameras. The primary contribution of this work is a degeneracy-aware, directional covariance-shaping strategy that replaces heuristic online covariance tuning with geometry-conditioned inflation. The core idea is to infer process noise from geometric corrections with a strict reliability check. We decompose the posterior-to-prior correction into a wheel-explainable subspace and an orthogonal remainder. The process covariance is anisotropically adjusted according to the observability of the ToF update: if the geometry is degenerate in non-kinematic directions, the remainder is suppressed as sensor noise; if the geometry is distinct, it is admitted as a valid kinematic violation. Additionally, we introduce an intra-frame feedback loop that retrospectively repairs the process covariance within the current frame to reduce structural lag during bursty disturbances. Extensive experiments on challenging scenarios demonstrate that TWO improves trajectory accuracy and robustness compared to static and adaptive Gaussian filtering, robustly navigating the trade-off between kinematic reliability and limited sensing capabilities.

Walking sequences reveal human-interpretable attributes related to body structure, posture, and motion. Gait attribute recognition provides a concrete route to fine-grained gait understanding by predicting these attributes. However, existing benchmarks provide limited fine-grained annotations, while classifier-based gait attribute recognition methods typically represent candidate classes through learned output weights rather than directly matching visual evidence with textual class representations. To address the annotation gap, we present RA-GAR, a large-scale, richly annotated multimodal benchmark for gait attribute recognition, containing 1,001 subjects, 242,729 walking sequences, and annotations for 15 static and motion gait attributes. We further propose Gait-VL, an attribute-centric vision-language alignment framework that recognizes gait attributes by matching attribute-specific multi-frame visual evidence with textual class prototypes. On the visual side, attribute-specific learnable queries aggregate RGB features across spatial locations and frames into per-attribute visual embeddings. On the language side, diverse descriptions expanded by a large language model are combined with context about relevant body regions to construct textual class prototypes. Beyond attribute recognition, we evaluate the transferability of the attribute-pretrained visual encoder by using it to replace generic upstream large vision models in identityoriented gait recognition pipelines; attribute pre-training uses no identity classification or metric-learning objectives. Experiments demonstrate state-of-the-art attribute recognition performance on RA-GAR and consistent improvements in downstream identity recognition across multiple pipelines and benchmarks, supporting the effectiveness and transferability of attribute-centric gait representation learning.

Test-time sampling promises to turn one action prediction into a set of distinct executable candidates that a verifier can choose among. Downstream fine-tuning works against that promise. A vision-language-action model trained to solve one task distribution concentrates its flow-matching decoder around the dominant demonstrated action, so repeated samples differ only marginally and the verifier receives many copies of one decision. We call this proposal contraction and identify it as a key constraint on test-time scaling: repeated samples can remain highly redundant as the candidate budget grows. Action Dynamic Sampling (ADS) alleviates the constraint on the proposal side, during training. At each training step, ADS integrates $K$ additional trajectories from the current policy under fresh base noise, scores each by the flow support along its path with a Hutchinson trace estimator, and corrects the policy's local action distribution toward this weighted set through a detached finite-set KL term. Training cost rises; deployment cost is unchanged. Across simulation and four physical-robot tasks, ADS raises proposal spread without the success penalty that Gaussian perturbation incurs, lifts oracle best-of-8 success on RoboCasa from 57.6% to 63.0%, improves fixed verifiers by 2.4 to 4.9 points in simulation, and raises TokenPB by up to 13.0 points on real-world manipulation.

This paper proposes a novel soft robot model named piece-wise constant 7D deformation (PC7) model that can represent expanding and contracting deformation of, for example, artificial muscle actuator. In the proposed model, we newly define the 7D deformation vector based on SIM(3) - 3D similarity transformation group - whereas a 6D deformation vector was defined for Cosserat rod based on SE(3) - the special Euclid group - in the piece-wise constant strain (PCS) model proposed in the previous study. The additional dimension of the 7D deformation vector represents expanding and contracting deformation, which comes from the scaling parameter in SIM(3). Common mathematical properties between SE(3) and SIM(3) allow us to directly extend the kinematics and dynamics of the PCS model to those of the PC7 model. In particular, we derive the dynamical equation of the PC7 model based on Euler-Poincare equation in a mathematically precise manner. We validate the proposed model through dynamical simulations assuming a thin cylinder structure or McKibben-type artificial muscle actuator.

Autoregressive robot video world models based on discrete visual tokens remain unreliable over long rollout horizons. Retaining the entire generated history does not necessarily improve prediction quality. We investigate a possible memory-role conflict in Full Context decoding: persistent scene information, recent visual-action states, and stale generated history compete for attention despite their different predictive roles. This perspective motivates ReCAP (Receding Context with Anchor Preservation), a training-free context-scheduling method for long-horizon robot video world models. ReCAP keeps the pretrained tokenizer, world model, and visual decoder frozen. It preserves the initial scene anchor and a contiguous window of recent complete visual-action blocks while evicting older generated history. The schedule preserves within-block structure and temporal order, requires no additional modules or parameter updates, and bounds active-context growth independently of rollout length. We evaluate ReCAP on RT-1, BridgeV2, LIBERO-90, and CALVIN using frozen autoregressive world models. Across evaluated environments and horizons, ReCAP improves visual fidelity, action-conditioned consistency, and robot structure preservation relative to Full Context. On BridgeV2, it reduces LPIPS by 63.3%. In the system benchmark at H=64, it reduces the logical KV footprint by 73.3% and p50 latency by 6.51%. Attention analyses show increased attention to retained recent states, while history interventions associate larger amounts of stale context with higher prediction error. These observations are consistent with the memory-role-conflict hypothesis, although they do not isolate it from context-length and position-handling effects. Our results support inference-time context scheduling as a practical approach to improving long-horizon robot video prediction and computational efficiency without retraining.

Distributed quantum computing distributes workloads that exceed a single processor's capacity across resource-constrained components. Circuit cutting enables this decomposition, but its effectiveness depends on partition choice. We present SACE, a Simulation-Aware Circuit Partitioning framework with Entanglement-Capability-Aware Screening. EP-Cost screens partitions using a parameter-dependent entanglement-capability proxy and native implementation cost. Under reconstruction constraints, finite-bond refinement selects the candidate that minimizes the worst-subcircuit discrepancy between normalized MPS simulations at bond dimensions $\chi$ and $2\chi$. Across 14 exhaustively enumerated small-scale candidate spaces, EP-Cost-guided genetic algorithm selections attain co-optimal native quasiprobability decomposition (QPD) overhead, with a median within-circuit Spearman correlation of 0.9965 between EP-Cost and QPD overhead. At matched cut counts, EP-Cost reduces QPD overhead by $25.20\times$ and $16.89\times$ for 8- and 9-qubit QFT circuits, respectively, and improves sampled reconstruction. On 12-qubit approximate QFT, deterministic reconstruction reduces Pauli-observable mean absolute error in all six low-bond conditions and total-variation distance in five. With five cuts and 7,776 QPD terms, SACE is best on all four reconstruction metrics in four of six conditions. At 24 qubits, SACE improves exact-local QFT accuracy in all 30 instance-bond-dimension conditions against each EP-Cost-only, Kernighan-Lin, and spectral baseline. The 50-qubit QFT comparisons remain favorable as reference convergence improves. Weighted QAOA circuits provide complementary, more bond-sensitive evidence. Refinement benefits persist across three search methods; larger candidate panels improve the optimized proxy without guaranteeing lower exact-local error.

Embedding models underpin modern job-candidate matching: resumes, job descriptions, and queries are encoded as dense vectors, and recommendations come from nearest-neighbor retrieval over that space. Any social bias in the embeddings therefore feeds directly into HR recommendations. We audit six multilingual sentence-embedding models (LaBSE- and E5-family, both public and custom fine-tuned) using Bias-in-Bios (occupational gender bias) and StereoSet (stereotypes across profession, race, gender, and religion). We treat statistical validity as a first-class concern: we identify a non-independence error in a naive significance test, where pooling non-independent pairwise similarities inflated a trivial 0.0046 mean difference to $p \approx 0$, and replace it with a profession-level paired Wilcoxon signed-rank test, individual-resampling bootstrap confidence intervals, and permutation tests for StereoSet. Under the corrected methodology, knowledge-distillation (KD) fine-tuning of E5-Base significantly increases measured bias on Bias-in-Bios ($\Delta = -0.177$, $p < 10^{-8}$; more negative means more bias), although this measure is sensitive to global embedding geometry. The KD result is corroborated by pro-stereotype patterns in the StereoSet profession category (uncorrected $p < 0.05$). Binary-classification fine-tuning of the same base model shows no statistically distinguishable change ($p = 0.218$). Fine-tuning strategy is thus an actionable design lever for practitioners building HR recommenders. We also report which benchmark categories are underpowered, and note that two domain-adapted models could not be included in significance testing due to a software-access constraint, which limits our conclusions.

To enable large language models to generate outputs that align with human preferences, optimization algorithms based on the comparison between chosen and rejected output pairs have been developed and been proven effective. Some recent endeavors have utilized certain linguistically-derived assumptions about the preference pairs to guide the design of more precise optimization directions. However, human preferences are complex and hard to measure by simplistic criteria, and it is important to incorporate broader and more sophisticated insights about preferred vs. not-preferred language. In this study, we adopt a well-established psycholinguistic theory, Uniform Information Density, to develop a novel preference optimization algorithm, UIDPO. The basic assumption is that humans prefer to produce language in which information is uniformly distributed, because uniformity induces less cognitive processing effort. To implement this idea into the algorithm, we derive a compound metric for information uniformity by combining the global standard deviation and the local second-order difference, and integrate it to the objective of preference optimization. We compare UIDPO with SimPO, a previous state-of-the-art method, obtaining comparable and sometimes better performance on AlpacaEval2, ArenaHard, and MT-Bench. Additional linguistic analysis on complement clauses also shows that UIDPO can better reflect the tendencies of keeping information uniform in natural language than baselines. Our results show that the subtle human preference over the information rate in language is a helpful source for achieving better human alignment for LLMs. The code repository is published here: https://github.com/CLCS-SUSTech/UIDPO.

During human-robot interaction, sight plays an important role. When the human sees a robot, they can observe the robot's actions and better predict the robot's future behaviors. Conversely, when the human cannot see the robot, they become uncertain about the robot's intent and plans. In this paper, we focus on the human's field of view. Existing human-aware planners either ignore this field of view and assume the human can observe the entire scene, or intentionally design their interactions so the robot is always visible to the human. By contrast, we enable robots to infer what the human has and has not seen during interaction, i.e., the human's field of view. We then apply this estimate as a vision-aware filter on the robot's actions. Given a nominal robot policy and a model of the interaction, we roll out the policy over a short horizon to see how the robot's state updates. We then reweight the actions based on how those actions impact the robot's predicted visibility: higher weights are assigned to actions that bring the robot into (or away from) the human's line of sight. In practice, our framework provides an online estimation and filtering process that repeatedly updates its understanding of the human's field of view and then locally perturbs robot actions based on the human's vision. We evaluate our vision-aware filter on robots that assist humans in collaborative cooking and coordinate with them in driving environments. Across simulations and a 12-participant user study, we find that our vision-aware filter changes both performance and preference: participants completed over three times as many subtasks with in-view assistance while cooking, and had fewer collisions with cars that purposely stayed out of their view while driving.

As frontier LLMs gain more capabilities, the cost to run them also goes up. At the same time, small language models (SLMs) also gain more capabilities but have largely been able to keep their cost stable, offering 5-40 times cheaper inference for the narrow, repetitive subtasks that agents decompose work into. However, the standing limiting factor is getting the correct behavior out of a small model in a fast, efficient, and economical way. Unfortunately, in most cases, zero-shot small models fail narrow production tasks. Furthermore, retraining one every time the behavior changes is too slow to iterate on. We propose doing both in sequence. First, develop the desired behavior in prompt space, on a frontier model, with immediate feedback from live traffic. Then, once the behavior converges, distill the recorded production behavior (not synthetic imitations of it) into an SLM through supervised fine-tuning (SFT). We show a proper use case of this method through a scheduling agent for home-care agencies that fans out a single yes/no safety decision per caregiver-shift pair on GPT-5.4-nano. A LoRA-tuned Qwen3-8B trained on 11.6k recorded production decisions reaches 0.913 agreement, while the teacher (GPT-5.4-nano), re-run on its own recorded prompts, agrees with itself only 72.7% of the time, pointing to the fact that the SLM may also be more consistent. Furthermore, when the two disagreed, a blinded professional scheduler sided with the student 48-29. Finally, we find that the SLM has approximately 2.5 times lower median latency and is a fraction of the cost. We believe this lifecycle of quick building, live feedback, and cheap inference is a blueprint for others building similar systems.

In-context learning (ICL) allows large language models to use demonstrations without updating their parameters. For interactive agents, demonstrations are trajectories of reasoning, actions, and observations that illustrate how to complete a task. Reusing demonstrations in new situations is challenging: directly inputting full trajectories consumes limited prompt space and constrains the overall effectiveness, while natural-language summaries require the agent to translate general guidance into specific action choices. Therefore, we study whether expert action choices can directly supervise a reusable scoring rule. We propose CB-Agent, which extends concept-based ICL to sequential action selection. We extract environment-action decisions from expert trajectories and fit a concept vector by maximum likelihood to model the conditional distribution over candidate actions given the context. This model combines the concept with frozen language-model features and priors. During ReAct interaction, the agent uses this distribution to select environment actions without including demonstrations in the action-scoring context. Across TextWorld, ALFWorld, and ScienceWorld, CB-Agent achieves the highest cross-benchmark average among the evaluated methods at all three tested model sizes and ranks first on TextWorld and ScienceWorld at each size. ALFWorld experiments support the learned concept's contribution beyond prior-only scoring and show further gains when combined with Reflexion. Code is available at https://anonymous.4open.science/r/CB-Agent-DFF9/.

Execution is often used as terminal feedback: it scores a generated program, but the resulting state is not used for another learning action. Learning after execution instead uses the execution result as an intermediate state that supports an additional learning interaction for the same policy. This creates a credit-assignment problem because the two actions share the same final objective but play different roles in the trajectory. Accordingly, action-specific credit assigns the initial action credit from the final trajectory outcome and the corrective action credit for improvement over the intermediate state. Image-to-TikZ serves as an executable visual testbed for this setting. On 945 held-out examples, learning after execution strengthens the shared policy under single-call inference and also produces a post-execution corrective capability. Single-call render success reaches 92.70%, compared with 88.25% for continued single-stage GRPO after the same number of additional training iterations. This corrective capability is not reproduced by continued single-stage training and depends on the intermediate state. Alternative credit definitions also induce different group-relative preferences over the same trajectories. Together, these results suggest that executable tasks contain useful learning structure beyond a terminal reward. Code is available at https://github.com/tianyufengcn/Learning-after-Execution.

Reinforcement learning with verifiable rewards improves the reasoning capabilities of large language models, but remains computationally expensive and sample inefficient. On-policy distillation (OPD) provides dense token-level supervision from a stronger teacher, yet standard OPD distills every token regardless of whether the teacher's preference actually improves task reward. Existing token-selection methods rely on proxies such as uncertainty or teacher–student divergence, which need not identify reward-improving supervision. We introduce Reward-Aligned Distillation (RAD), a bilevel framework that weights distillation signals based on their predicted effect on the student's post-update reward. RAD differentiates through a temporary optimizer step to score each token's supervision, then updates the token weights by projected ascent from the initial token weights. The resulting weights can strengthen, suppress, or reverse individual distillation-gradient contributions. At a matched training-response budget, RAD improves mean accuracy on six competition-math benchmarks over the strongest baselines by 6.5 percentage points for Qwen3-1.7B (non-thinking) and 2.5 points for Qwen3-4B-Instruct-2507. RAD-Simple, a simplified variant, achieves comparable math accuracy at the same training-response budget. On LiveCodeBench-v6, RAD reaches 54.69% and 58.67% pass@1 with Qwen3-1.7B (non-thinking) and OLMo3-7B-Instruct, respectively, exceeding the strongest baselines by 8.7 and 6.4 points. These results support reward-guided weighting of teacher supervision across model families and tasks.

Large Language Model (LLM) inference is driving a growing share of datacenter carbon emissions. Although carbon-aware scheduling can reduce these emissions in geo-distributed datacenters by exploiting the spatiotemporal variations in grid carbon intensity, the latency-sensitive nature of LLM serving constrains this design space. While prior studies investigate the carbon savings achievable through spatial scheduling, they oversimplify the dynamics of modern inference engines, such as queueing and continuous batching, which strongly shape inference latency. We address this issue by designing and evaluating a latency-centric, geo-distributed inference simulator, calibrated with hardware-grounded latency and power measurements, that reproduces these dynamics and enables the study of carbon savings under strict service-level objectives (SLOs). We then formulate a series of SLO- and carbon-aware schedulers of increasing sophistication, from per-request heuristics to batch-level optimization: (i) a greedy heuristic, (ii) an integer linear programming (ILP)-based scheduler jointly assigning batches of requests, (iii) an iterative ILP-based solution projecting its own assignments back into the load estimates to capture load coupling, and (iv) a facility mixed-integer linear programming (MILP) scheduler pricing the energy of batched requests collectively rather than in isolation. Across three LLMs under MLPerf SLOs and varying system loads, the MILP-based option offers the best carbon-SLO balance, reducing carbon emissions by up to 60% relative to the round-robin policy, with an average SLO violation of 3%. Our characterization further shows that savings grow with the spread of carbon intensity across datacenters and shrink as system load increases, while SLO attainment depends on how frequently schedulers receive datacenters' load updates.

Large language models (LLMs) support social simulation, but faithfully reproducing real users requires grounding both structured platform decisions and language in sparse, noisy behavioral evidence. We introduce PsyMemBehav, which maps psychology-informed profiles and retrieved memories to target-first platform behavior through two-level identity-to-action reasoning. To support this approach with comprehensive behavioral traces, we construct a dataset PsyMemBehav-bench from 5.57M observed behaviors of 5,862 real users (approximately 950 per user) across Reddit, Bluesky, and Twitter. Each recorded engagement is modeled as target selection followed by Comment, Like, or Repost and optional response generation. PsyMemBehav combines 41-dimensional user descriptors, top-5 BM25 memories, reverse-CoT supervision, and rubric-guided DPO over complete reasoning-action trajectories. Experiments show that PsyMemBehav 27B ranks first in 11 of 12 reported individual-level metrics, improving average response quality over GPT-5.4 from 54.28 to 59.68 (9.95% relative), and obtains the lowest group-level MAE (0.081) and MSE (0.009) among simulated systems. These results support evidence-grounded user modeling for observed-thread continuation; the benchmark remains conditional on recorded engagement and does not model no-action decisions.

In autoregressive video generation, retaining all historical key-value (KV) states increases memory usage and per-step attention costs as videos grow longer. A fixed cache budget bounds these costs but raises the challenge of preserving historical content that supports consistent, high-quality generation. Existing methods control cache size through history selection or consolidation, yet deciding what to retain and how to limit the loss of useful content remains a shared challenge. Our analysis shows that spatial queries depend on different historical content and that higher current attention coverage does not necessarily yield better long-video quality. We introduce FrameFold, a training-free KV cache consolidation method that compares the content retrieved by the same recent historical queries before and after merging to guide both representative construction and merge selection. Exploiting temporal redundancy between adjacent historical entries, FrameFold merges their KV states into shared representatives while retaining complete spatial grids. It fits value-merging coefficients in closed form by matching conditional attention responses, then selects the merge with the lowest remaining response distortion. Under matched KV cache budgets, experiments on Self Forcing and Causal Forcing at 30 and 60 seconds and on LongLive at 120 seconds demonstrate improved consistency and visual quality. Continuous eight-minute Self Forcing rollouts further show less quality degradation than the baselines.

This paper develops strong Gaussian approximations (KMT type) for constant learning rate stochastic gradient descent (SGD). We work with heavy tailed noise and assume that the number of iterates grows and the learning rate shrinks simultaneously. Our work constructs simple Gaussian processes on the same space and bounds, in probability, the maximum difference between the SGD errors and their respective limits. While there exist central limit theorems for constant rate SGD centered at the optimum, these only provide individual iterate inference. By contrast, our quenched KMT approximations allow for finite sample, simultaneous, whole-trajectory-level inference and imply stochastic process convergence (Functional Central Limit Theorems (FCLTs)). For accumulated errors centered at the true optimum, we first obtain a KMT-type approximation bound when the learning rate scales as $\gamma_n=\frac{c}{\sqrt n}+o(n^{-1/2})$ for a constant $c \in \mathbb{R}^+$ and $n$ iterations. The result reveals a trade off between controlling the bias of SGD and the initialization effect, with no learning rate scaling generally making both contributions vanish in the accumulated error limit. To eliminate these contributions, we consider a Burn in scheme and Richardson-Romberg extrapolation and prove respective KMT and FCLT theorems. We also prove a finite-horizon KMT-type approximation and FCLT for the rescaled iterate trajectory to a stationary Ornstein–Uhlenbeck (OU) process. We consider applications in risk control and simultaneous inference and perform simulations to illustrate our results.

Dynamic flux balance analysis (dFBA) and its community, spatial and ensemble variants predict microbial metabolism by solving one linear program (LP) per species and time step, often 10^6 to 10^8 LPs per study. These LPs share their constraint matrix and objective and differ only in their bounds, yet general-purpose solvers treat each one independently, and GPU first-order solvers struggle with the degenerate, badly scaled matrices of genome-scale models. Moreover, metabolic optima are usually faces rather than vertices, so solvers return arbitrary flux vectors and dFBA predictions depend on the solver used. We present manylp, a batched solver for such bounds-parametric LP families that returns certified vertex solutions. Building on the bunching technique of stochastic programming, manylp caches optimal bases and certifies whole batches with one dense matrix product per basis on CPUs or GPUs; LPs outside the cached critical regions are re-solved by a warm-started lexicographic dual simplex, and new bases are propagated across the batch. Each LP is certified optimal and unique under a lexicographic flux rule, or infeasible, in independent double-precision arithmetic. On a genome-scale dFBA workload, manylp certified 47,889 LPs per second on a CPU, 12 to 33 times faster than HiGHS, Gurobi, FICO Xpress, GLPK and OR-Tools (32 processes), and reached 682,000 LPs per second on one GPU for community ensembles, needing 76 simplex solves for 94.6 million LPs. Integrated into the muODE simulator, it accelerated single trajectories 84-fold and 16-member ensembles 635-fold, and made adaptive Runge-Kutta integration affordable, reducing biomass errors 83-fold. Integrating 49 equally optimal flux policies in one batched run showed that alternative optima change predicted acetate 4.2-fold and shift community composition. manylp is open source: https://github.com/cumbof/manylp.

Belief propagation in social networks depends on how trust and distrust jointly shape interpersonal influence. We develop a network model that treats trust and distrust as distinct but statistically coupled quantities. Their joint values determine the net trust and uncertainty of each relationship, and a neighbor can influence an agent only when both threshold conditions are satisfied. We show that the correlation between trust and distrust reshapes the transition between consensus and fragmentation by changing the availability of influential neighbors. Network connectivity also affects this transition, whereas degree heterogeneity has a more limited effect when networks have comparable mean degree. Moreover, the availability of influential neighbors alone does not determine the final consensus: correlations between relational properties and network degree can select agents with different degree compositions, altering the direction and magnitude of collective belief. Finally, we introduce a comfort index that quantifies the expected availability of influential neighbors and analytically links the joint trust-distrust structure to collective dynamics. Our results establish a mechanistic connection between relational structure, effective social influence, and emergent collective belief states.