Sección

Investigación

Papers y preprints de los feeds de investigación que seguimos.

6265documentos

arXiv AI Papersanteayer

ARE-bench: Benchmarking Agents' Research Ability in Real-World Researcher Sessions and Workspaces

Research agents achieve increasingly strong benchmark scores, yet these scores do not necessarily translate into reliable delivery in ongoing research projects. Many research-agent benchmarks emphasize standalone tasks and relatively self-contained workflows in clean or curated workspaces. However, researchers ask agents to complete one step within an ongoing project, where requests depend on earlier session decisions and workspaces contain unfinished files, outdated versions, and failed attempts. We introduce ARE-bench to evaluate whether research agents advance ongoing projects while respecting prior decisions and existing workspace state, grounded in real session context and in-progress workspaces. ARE-bench includes 75 tasks built on agent sessions and full workspaces from recruited researchers' own projects, spanning seven research stages, 33 requirements, and 150 sub-requirements, with 1,173 binary checklist items in total. Each task is anchored at a real user request, preserves its requirements, and summarizes the earlier requests still in effect as background. The complete workspace is replayed to that boundary without cleanup, preserving artifacts from earlier turns. Each checklist is compiled from an intent-aligned rubric and credits any valid solution. Across seven frontier models, the strongest passes every checklist item on only 66.2% of tasks, and Design and Execution rank among the hardest stages for every model. Three models pass about 69% of tasks when only delivery contracts and general norms are checked, but just 28.4% to 40.0% once historical and material or workspace requirements are also checked. Evaluating delivery without these contextual requirements can therefore overestimate research agents and mask differences between them.

en

arXiv AI Papersanteayer

The Unused Input: The Initial State as a Conditioning Interface for Recurrent TSFMs

Time series foundation models (TSFMs) forecast unseen series without task-specific training. Increasingly they are built on recurrent backbones, whose constant per-step cost makes long-context inference cheap, and which, unlike transformers, carry a state. A recurrent backbone computes its prediction from two inputs: the observed context window and the initial recurrent state. In practice only the first is utilised and the initial state is unused by convention. We ask whether the initial state can be utilised in recurrent TSFMs. Using the xLSTM-based zero-shot forecaster TiRex as a case study, we freeze the backbone and train only a small head that writes a context vector into the initial state, so that any change in accuracy is attributable to the injected state rather than to adapted weights. A spectral summary of the input window, learned once and applied unchanged to every series, lowers mean GiftEval MASE by 0.175%, with every seed improving, which neither a parameter-matched LoRA adapter nor a learnable static state reproduces. An embedding of text attached to each series, evaluated on six MoTime datasets, reveals two roles for the state. As a per-dataset adaptation surface, a static state lowers MASE by 0.04-6.96% of the zero-shot baseline (8-70% of the gain of full fine-tuning). As a channel for exogenous information, the text-conditioned state lowers it by a further 0.15-0.41% of the zero-shot baseline on four of six datasets and, combined with a LoRA adapter, exceeds full fine-tuning, which has no access to the text, on one of them. In TiRex, the effect concentrates in the mLSTM blocks and is strongest at short context. The results are preliminary and the improvements small, but they point to the initial state as a direction for further research.

en

arXiv AI Papersanteayer

EngineAgent: Agentizing On-Device LLM Inference Engines with Inference Stack Hierarchy Awareness and Hardware Abilities

Rapidly evolving LLM architectures and heterogeneous mobile hardware make supporting new models in on-device inference engines a recurring and costly system challenge. Each model may require coordinated changes across conversion, graph construction, operators, and hardware backends. Existing coding agents often miss these cross-layer dependencies and struggle to diagnose failures across host--device execution paths. We present EngineAgent, an agentic system for automating on-device LLM model enablement. Its key insight is to separate reusable engine and hardware knowledge from model-specific implementation. Before enablement, EngineAgent constructs a hierarchy-aware engine index that captures engine abstraction layers and cross-layer dependencies, and pre-validates backend-specific build, execution, and debugging workflows using already supported models. For each new model, a builder uses this knowledge to coordinate cross-layer changes, while an auditor checks implementation completeness and drives iterative debugging through device-side validation. Across 12 tasks spanning three mobile LLMs and four engine--backend pairs on CPUs, GPUs, and NPUs, EngineAgent achieves 75% end-to-end success with either of two underlying LLMs, versus 33.3% for the strongest coding-agent baseline. Success requires compilation, on-device execution, and output-logit agreement with reference implementations. EngineAgent also reduces average per-task time and model-API cost relative to baselines using the same underlying LLMs.

en

arXiv AI Papersanteayer

CLADE: Category-Level Answering from Deficient Evidence in Retrieval-Augmented Language Models

Retrieval-augmented systems often retrieve passages that are accurate and relevant but state the answer at a coarser level of granularity than the question requires. For example, a question may ask which specific disease a drug treats (familial hyperlipidemia), while the retrieved passage mentions the broader disease family (hyperlipidemia). Prior work on context sufficiency treats such coarse passages simply as insufficient. We argue that they form a distinct case, which we call resolvable insufficiency: text alone cannot settle the fine-grained answer, but the broader category it mentions maps to a short list of candidate answers in a domain ontology. When language models read these coarse, family-only passages, they exhibit a phenomenon we call Clade (Category-Level Answering from Deficient Evidence): they answer instead of abstaining or searching further. On biomedical questions from PrimeKG, family-only passages reduce abstention by 13.8 points compared to uninformative passages, and the resulting guesses are wrong about half the time. A counterfactual test confirms the cause: replacing an exact disease name with its broader family name lowers abstention by 17.4 points compared to deleting the sentence. Simple fixes like confidence filtering, scaling to a larger model, and explicitly instructing the model to abstain cannot resolve Clade, and this replicates on ecological species interactions in GloBI. Finally, we test whether an ontology can turn coarse passages into exact answers. Given the correct family, an ontology list includes the true answer in 98.5% of cases, but string matching finds that family on the page for only 19.8%. Letting the model choose the family raises this to 30.0% and end-to-end recovery from 10.0% to 26.8%. Overall, our results show when language models should stop guessing and consult an ontology, and highlight which steps of that process remain unsolved.

en

arXiv AI Papersanteayer

AI Isn't Getting Dumber, We're Getting Pickier: Rising Standards in Human Evaluation of Large Language Models

Users often complain that AI assistants such as ChatGPT are getting worse. Some versions do change, but human evaluation leaves a second possibility untested: the people judging may change. We separate the two on Chatbot Arena, a public platform where people compare answers from two anonymous large language models (LLMs) and pick the better one or say both are bad. We study 1,260,967 blind votes cast from May 2023 to August 2024 on pairs of models whose versions are fixed: dated vendor snapshots and models with fixed open weights. For the same pair of unchanged models, the share of votes saying "both are bad" rises by 1.71 percentage points a month on a mean of 17.1%, and by 1.91 points a month in 2024. In fourteen pairs observed often in both 2023 and 2024, it rose in every one, from 15.5% to 23.3%. The ranking does not trend: in 2024, the stronger model's share of decisive votes is flat. The rise holds within the same voter and on the same pair, and across arrival cohorts from their first recorded vote. It holds with prompt controls, in every major language, for dated vendor snapshots, open-weight models and models with stable answer length, and between interface changes. The drift is shared across models: month effects estimated from other models' pairs remove it from each model's own pairs. Platform-wide, the average model would be expected to score 78% against the 2023 average, yet the share of "both are bad" votes ended near where it began. The pattern fits a standard that follows the category's progress, so a product that stands still is judged as declining. Absolute verdicts on AI drifted while relative comparisons held steady, an evaluation drift that matters for LLM leaderboards and for the human feedback used to train models.

en

arXiv AI Papersanteayer

Forward-Value Discretization: Faster Stochastic Diffusion Sampling Without High-Order Approximation

Stochastic diffusion samplers often require substantially more neural function evaluations (NFEs) than deterministic samplers to achieve comparable sample quality. Although higher-order discretization accelerates deterministic sampling, its benefits for stochastic sampling remains unclear. We establish matching upper and worst-case lower bounds showing that stochastic samplers with higher-order approximations remain first-order convergent, even with exact score functions. Thus, increasing the approximation order does not generally improve the convergence order. Motivated by this limitation, we introduce forward-value discretization, a training-free framework that evaluates the data/noise prediction model at a lookahead estimate constructed from previous predictions. For deterministic sampling, the resulting Forward DPMSolver (F-DPMSolver), improves sample quality at low NFE budgets through its DDIM-based variant, despite remaining first-order convergent. For stochastic sampling, we develop Forward DDPM (F-DDPM), which evaluates the model at a cleaner lookahead point adjusted for subsequent noise injection. This estimate reuses both the previous model prediction and the previously injected noise. Both variants admit first-order convergence guarantees with one model evaluation per iteration. Across CIFAR-10, ImageNet, FFHQ, and LSUN, F-DDPM consistently achieves lower FID than the evaluated stochastic baselines at matched NFE budgets. With 1.5 times as many NFEs as deterministic baselines, it achieves competitive sample quality, narrowing the practical efficiency gap between stochastic and deterministic sampling. Code is available at https://github.com/Na-Li66/F-DDPM.

en

arXiv AI Papersanteayer

When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents

LLM agents increasingly operate in long-horizon, verifier-rich environments, but current self-improvement mechanisms often write episode-derived lessons directly into memory, making local success an unreliable signal for reusable and safe knowledge. This paper addresses the core question of when an experience-derived behavioral update should be admitted into persistent agent state. We propose Falsification-Gated Self-Evolution (FGSE), a verifier-grounded framework that represents each candidate lesson as a structured hypothesis with an explicit precondition, behavior change, expected effect, and verifier. FGSE applies the hypothesis only in a temporary state, tests it on target-transfer, falsification, and archived-regression probes, and then commits, refines, or rejects it according to measured gain and risk. Across web, app, tool-use, and long-term memory benchmarks, FGSE achieves strong task performance, including 39.8 WebArena success rate, 78.8 AppWorld average completion, and 0.724/0.487 Pass^1 on τ-Retail/τ-Airline, while reducing falsification failures to 5.0%, regression damage to 1.3%, and harmful committed updates to 3.6%. These results suggest that testing memory updates as hypotheses supports more reliable long-term adaptation.

en

arXiv AI Papersanteayer

Point-in-Time Encoders for Earnings Calls

Text encoders turn earnings calls into vectors for search, for language models that answer from retrieved passages, and for forecasting models. A live retriever can use an encoder trained once on all available text, as finance retrievers are, but a backtest needs an encoder that read no text written after the date of each vector. To address this look-ahead bias, from 412,248 calls we train vintages under four pretraining recipes, with yearly cutoffs from 2010 to 2024, and each vintage reads only text dated on or before its cutoff. We then fine-tune each vintage on the question-answer pairs of the calls themselves. The vintages we train on calls alone, with 33.8M parameters, retrieve answers on held-out calls better than the general retrievers we tested in every test quarter. Their vectors also raise by about half the rank correlation between next-day returns and forecasts built from past returns and firm size. When each vintage continues the previous one, a later vintage retrieves better on earlier quarters than the vintage available at the time, and a point-in-time language model that answers from the passages the later vintage retrieves inherits the gain. The 2024 vintage of our families also forecasts earnings surprises better, although the interval of this gain includes zero on calls that no vintage read. It forecasts returns no better, so point-in-time vintages matter most for retrieval and little for forecasts of returns.

en

arXiv AI Papersanteayer

Auditing the Judge: Internal Signals for LLM Reasoning Assurance

Language-model systems use self-judgment to decide which answers to release, review, or recompute. Because intervention is costly, an assurance score must rank errors well at the available budget, not merely over the full ranking. We compare the judge's continuous verdict-token score with a supervised correctness probe trained on attention-head activations from the same prompt. Exact identities link fixed-budget error recall to residual risk and show that AUROC averages recall over all budgets. Across three open 7-12B models on GSM8K and MATH, the probe finds 1.3-22.5 additional percentage points of errors at 5% review in all six model-dataset settings (median 7.4 points). At 10% routing, probe-guided allocation improves accuracy over equal-cost random allocation in both GSM8K experiments, but beats judge-guided routing only for the weaker judge. Cross-dataset transfer lowers probe AUROC for every model in both directions relative to target-domain training. Because dataset, weight precision, and generation budget change together, this measures a joint protocol shift rather than a dataset-only effect. Internal signals can improve supervised, in-domain budget allocation, but are not domain-general confidence estimators.

en

arXiv AI Papersanteayer

Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents

External memory enables LLM agents to improve from experience by distilling trajectories of past tasks into concise lessons for new tasks. However, memory can also lead to memory-induced degradation, where an agent fails on tasks it would solve without memory, often because it follows a lesson that is incorrect, or misapplies or misreads a valid one. This is because lessons are inferred from trajectories labeled only with task success or failure, yet are typically provided without the trajectories they came from, leaving the agent no way to check them. To address this, we introduce WikiTrace, an evidence-linked memory system designed to let agents check each lesson against its original trajectory. Specifically, motivated by the finding that a trajectory often still holds the evidence needed to catch an erroneous lesson, WikiTrace stores each lesson as a concise insight linked to a summary of why its trajectory succeeded or failed and to the raw trajectory itself. Then, on a new task, the agent starts from all insights and, whenever one lacks details needed to act, conflicts with another, or comes from a failed task, descends through its links with terminal tools to fill in missing details or correct the insight before acting. We validate WikiTrace on five agentic benchmarks, showing that it consistently outperforms existing memory approaches and a baseline given the same insights alone, while also reducing memory-induced degradation. Our code is available at https://github.com/yukyeongleee/wikitrace

en

arXiv AI Papersanteayer

Generalizing Causal Decision-Making: Markov Equivalence, Transportability and Counterfactuals

Structural causal bandits use causal knowledge to guide exploration when reward mechanisms are unknown. This paper extends the approach to partially known causal structure, information from other environments, and realizable counterfactual actions. First, we develop graphical criteria for minimal intervention sets under Markov equivalence and characterize possibly-optimal minimal intervention sets for maximal ancestral graphs. A search algorithm finds sets that are possibly optimal in at least one graph represented by a partial ancestral graph. Second, we derive target reward bounds from source distributions through partial transportability, accounting for non-manipulable variables. Dominance relations provide further bounds for action pruning and restricted exploration, with a regret guarantee under valid prior bounds. Third, we formulate counterfactual structural causal bandits over actions supported by specified experimental capabilities. Graphical characterizations of minimality and possible optimality yield algorithms that remove equivalent or dominated choices while preserving an optimal action. Experiments with specified structural causal models examine how these reductions and transferred bounds affect learning. Together, these results clarify how causal knowledge and experimental capabilities determine which actions merit exploration and guide learning within the retained action space.

en

arXiv AI Papersanteayer

Where Do Vision-Language Models Get Left and Right? Position-Code Counterfactuals in Qwen2.5-VL, Qwen3-VL and InternVL3

A vision-language model may take left-right direction from the orientation of image content within patches, its vision transformer's (ViT) position code, or token positions and order in its language model (LLM). Ablations show the ViT code is necessary; we test sufficiency by mirroring only a position code. In Qwen2.5-VL-7B, Qwen3-VL-4B and InternVL3-2B, mirroring ViT width positions reproduces the flipped-image answer on 84.3-93.2% of CV-Bench left-right questions (Qwen2.5-VL-3B: 87.5%, partly a left fallback). With two options, this also counts disruption; a half-grid shift flips most answers too. With the ViT code intact, LLM-side id and token-order mirrors change few left-right answers in all three series (a weak id test in Qwen2.5-VL). In Qwen2.5-VL-7B the effect sits in the global-attention ViT blocks, mostly the last. On What'sUp, flipping content and token order, not the position code, keeps more on/under than left-right answers at the flipped-image answer. Post hoc analyses qualify these results. On CV-Bench items with located objects, random permutations flip 18.9-31.0 points more answers when reversing the objects' order than when keeping it. On What'sUp, in Qwen2.5-VL and Qwen3-VL, the share of the left-right changes made by random permutations that land on the mirror relation is mostly below chance. Counted as returns to the original answer, the What'sUp contrast of support with left-right relations is clear only in InternVL3-2B. Most remaining errors give the same relation whichever object is asked about, suggesting that repair may target relation binding, object localization or use of the reference object.

en

arXiv AI Papersanteayer

HandsOnAid: Building and Evaluating an Offline First-Aid Assistant for Emergency Response Scenarios

Emergencies can occur where there is no phone signal and no ambulance nearby. A small language model on a phone can be helpful in such situations, but its advice needs to be safe and accurate. We built HandsOnAid, a 4-bit Gemma 2B model fine-tuned with QLoRA on FirstAidQA dataset, a corpus of 5,550 first-aid question-answer pairs. We compared seven configurations, including the base, 4-bit and 8-bit fine-tuned models, and versions with a length floor, safety gate, or retrieval. We used a 41-pair question-answer bank with medically reviewed references that followed ANZCOR guidelines, and a strict 0-5 score rubric, with twelve safety overrides. Four frontier LLM judges (GPT-5.6 Sol, Claude Opus 5, DeepSeek V4 Pro, GLM-5.3) blindly scored every answer, with 45 hidden controls used to check judge reliability. Fine-tuning improved the mean score by +1.08 over the base model to 2.34. The best configuration scored 2.51 out of 5 which was the fine-tuned 4-bit model with retrieval. A medical professional also conducted a blind evaluation of 40 answers across the base, the 4-bit fine-tuned, and the retrieval configurations, and the scores were similar to the LLM judges' average scores. A keyword audit found limited coverage of several details of ANZCOR guidelines in the training corpus, that are required by the rubric, which suggests that missing information in the training data could be a cause of the low scores. On a smartphone device (Pixel 7 Pro) with six CPU threads and no network connection, the 4-bit fine-tuned model produced its first displayable text in 6.86 seconds and a complete answer in 14.88 seconds on average over ten questions.

en

arXiv AI Papersanteayer

Can KV Cache Be Reused Across Model Updates? A DeepSeek-V4-Flash Case Study

Reusing key--value (KV) caches across language model updates could avoid repeated computation of shared context, but changes in model weights raise a fundamental question about what these caches preserve. We investigate direct cache reuse across four model updates, with a detailed study of DeepSeek-V4-Flash to Flash-0731. Our central finding is that caches can retain useful task information without faithfully reproducing the updated model's behavior. Retrieval accuracy varies substantially across updates, ranging from near-native performance to large losses. With source caches already available, reuse achieves median time-to-first-token speedups of 2.10--5.08on three longer-prefix QA benchmarks. To characterize the behavioral differences behind these outcomes, we introduce Functional Cache Divergence, which compares next-token distributions under reused and native caches. Direct KV cache reuse better preserves native model behavior in natural-text continuation, whereas compatibility degrades with increasing prefix length in single-token retrieval tasks. Answer-margin analyses help explain how correct retrieval decisions survive substantial internal state changes. Yet even agreement on the first generated token does not ensure agreement in answer content or stopping behavior. These findings establish task utility and fidelity to native inference as distinct criteria for evaluating cross-version KV cache reuse.

en

arXiv AI Papersanteayer

Structured Confidence Perturbation via Fisher Geometry for Adversarial Training

Adversarial training is often plagued by "pseudo-attacks", perturbations that significantly increase the loss value but fail to truly alter the model's confidence structure. To address this, we propose two complementary approaches to directly attack and defend against confidence structure deformation. First, building upon the traditional loss maximization perspective, we introduce a hybrid objective that integrates cross-entropy and KL divergence, encouraging adversarial examples to simultaneously disrupt prediction accuracy and distribution consistency. Second, we geometrically reconstruct adversarial training by generating perturbations that maximize the distribution distance on the confidence manifold. Theoretical analysis shows that the optimal perturbation direction aligns with the principal eigenvectors of the Fisher information matrix. We also propose an efficient algorithm to avoid high-dimensional matrix computations. Experiments on multiple datasets and model architectures demonstrate that our method outperforms existing methods in adversarial robustness and robust transferability, particularly excelling in structural perturbations within high-confidence regions.

en

arXiv AI Papersanteayer

Thermodynamics of words

We introduce the word thermalization hypothesis according to which in a literary work there are effective interactions between words that drive them to thermalization where their word frequency distribution is the same as for the phenomenon of Rayleigh-Jeans (RJ) thermalization and condensation. In this concept, the different words play a role of classical oscillators (particles) and their word frequencies, or related tokens, are analogous to certain filling factors over energy state levels. The thermalization process is characterized by the conservation of total system energy, or total number of words/tokens, and total probability norm, or total number of different words. We apply this approach to analyze 16 literary works, including Bible, complete works of Shakespeare, and data from American National Corpus. We show that for these cases the effective system energy is very low and thus there is the RJ condensation phenomenon with a big fraction of total probability concentrated at the lowest energy states corresponding to a big fraction of rare words appearing only once or twice in the whole text. Such a RJ condensation is observed for light propagation in multimode optical fibers and other physical systems. We relate the thermalization of words to the recently proposed wealth thermalization for households in world countries. The economy concepts of Lorenz and Pareto curves, and also Zipf curves, are applied to the word frequency distributions of the selected 16 cases and it is shown that the thermalization theory gives a strikingly good description of real data.

en

arXiv AI Papersanteayer

MISO-BR: An Open Corpus for Misogyny Detection in Brazilian Portuguese

Misogyny is among the most prevalent forms of online hate speech directed at Brazilian users, yet open and auditable resources dedicated to it in Brazilian Portuguese remain scarce. We introduce MISO-BR, an open, documented and versioned corpus of 896 balanced social media texts annotated for binary misogyny, released with its construction protocol and reproducible baselines. The corpus combines lexicon- and target-guided scraping from X with items imported from OLID-BR, a large language model triage stage whose prompt and decoding parameters are released, and manual annotation by three annotators following a written guide. Inter-annotator reliability reaches a Fleiss' kappa of 0.788 over 816 items. Under a single fixed protocol we compare four classical classifiers (Multinomial and Bernoulli Naive Bayes, a linear SVM and a random forest) with fine-tuned BERTimbau Base and Large. The classical models cluster between 0.651 and 0.676 macro-F1 and are precision-oriented but recall-poor. The transformers reach 0.833 macro-F1 and accuracy (Wilson 95% CI for accuracy [0.77, 0.88]), a gap of roughly 16 points, while Base and Large are indistinguishable on the test set despite a 2.6x difference in training cost. We argue that the value of the resource lies less in its headline scores than in the auditability of its pipeline, and we state three limits that users must reckon with: the artificial class balance, whose effect on precision under realistic prevalence we quantify; the association between provenance and label, which the released source field makes measurable; and the selection bias induced by lexicon-guided sampling and automatic triage. The corpus, annotation guide, models and experimental scripts are publicly available.

en

arXiv AI Papersanteayer

ScaleBrain: A Scale-Consistent Framework for EEG Foundation Models

Foundation models pre-trained on large unlabeled electroencephalography (EEG) corpora have emerged as a transformative paradigm for Brain-Computer Interface (BCI) applications. Recent EEG foundation models have shown that multi-scale representations derived from fixed-length patches can improve EEG representation learning. Existing frameworks commonly construct temporal scales through feature-level operations. However, these operations leave the correspondence between learned features and scale-specific characteristics of the original EEG signals implicit. Moreover, when the self-supervised objective is not aligned with the multi-scale input representations, information at some temporal scales may receive insufficient supervision during pre-training. In light of this, we propose ScaleBrain, a scale-consistent pre-training framework that explicitly coordinates multi-scale representations of EEG from both the model input and optimization objective sides. Specifically, ScaleBrain introduces multi-scale EEG tokenization to represent the same EEG segment through analysis windows with different temporal spans and shared temporal anchors, and pairs these representations with scale-specific reconstruction to align with the multi-scale input structure and provide supervision across temporal resolutions during self-supervised learning. We further develop an axial spatio-temporal attention mechanism to contextualize the fused multi-scale representation along spatial and temporal dimensions. Extensive evaluations on six public BCI benchmarks against 12 representative baselines show that ScaleBrain achieves the highest mean Balanced Accuracy on five datasets and ranks third on the remaining one, with only 5.0M parameters, demonstrating strong performance and favorable parameter efficiency.

en

arXiv AI Papersanteayer

Few-Step Streaming Video Generation via Condition Shifting

Few-step sampling reduces the cost of streaming video generation. Distribution Matching Distillation (DMD) supervises few-step generators with a separate fake-score network, which adds training memory and computation and must track the evolving generator. We extend APEX to streaming video by using condition shifting to estimate the velocity field of the generated distribution within the generator. The original text condition selects generation, while a shifted condition selects fake estimation using the same parameters. This removes the independent fake network while retaining S2-Avatar's frozen reference teacher, Self-Replay Forcing, and perceptual regularization. We evaluate APEX on 180-second videos using three backbone evaluations per block and a one-step spatial Refiner, with a 70-case human study and a 320-case StreamAV-Bench comparison. Compared with S2-Avatar, APEX achieves higher video instruction fulfillment on the Progressive track (3.856 versus 3.831) and higher visual aesthetics across both tracks (0.641898 versus 0.632606). Human raters favor APEX most strongly on instruction following, with +30.0% net preference, followed by consistency and video quality at +4.3% each. Motion quality favors S2-Avatar, audiovisual synchronization is predominantly tied, and usability is unchanged. Separate measurements under matched training settings show 4.4% lower peak allocated GPU memory per GPU and 9.2% shorter mean iteration time than DMD.

en

arXiv AI Papersanteayer

Program-Aided Patchable Skill Evolution

Agent skills have emerged as an effective means of enabling LLM agents to solve complex, interactive tasks by encoding reusable knowledge without modifying model parameters. Recent approaches further improve skill adaptation by distilling reusable experience from agent execution trajectories and evolving skills based on downstream task feedback. However, existing methods primarily represent skills as natural language documents, intertwining procedural logic with descriptive knowledge and making skills difficult to maintain and evolve. To address this limitation, we propose PASE (Program-Aided Patchable Skill Evolution), a program-aided framework that iteratively evolves skills by associating them with executable programs. Unlike prior approaches that encode trajectory-derived experience primarily in natural language, PASE separates reusable procedural logic into editable programs, enabling targeted updates to executable components while reducing the need to repeatedly revise natural language skills. Extensive experiments across diverse models and benchmarks demonstrate that PASE consistently outperforms all baselines, by achieving over 2.8% improvements. Further analyses show that PASE exhibits greater evolutionary stability and generalisation capability, highlighting the critical role of programmatic knowledge in skill evolution. We will release all data and code via GitHub.

en