FabriMAE I Trust Myself? Self-Evaluating VLA Action Generation with Markov Attention Entropy

Vision-Language-Action models (VLAs) integrate visual perception, language instruction, and action generation into end-to-end policies across heterogeneous architectures. However, enabling VLAs to self-evaluate their action generation reliability without external supervision remains a major challenge. Existing methods either rely on expert annotations or estimate uncertainty only from output statistics, largely ignoring internal signals. In this work, we observe that internal visual modality entropy exhibits consistent distinctions between successful and failed tasks across heterogeneous VLAs. Although VLAs' architectures differ in their action generation, we show that they share a common latent action generation abstraction evolving under visual perception, language instruction, and state input, which we formulate as a Conditional Generative Markov Chain. Based on this formulation, we propose MAE (Markov Attention Entropy), a self-evaluation framework that directly converts internal attention signals into architecture-aware reliability scores, and introduce LIBERO-Reflect, a 4,000-episode benchmark combining 2,000 standard episodes and 2,000 challenging episodes across four subsets. Extensive experiments across heterogeneous VLA architectures and diverse scenarios show that MAE consistently outperforms state-of-the-art baselines on AUPR, AUROC, and FPR@95. We further instantiate FabriMAE for verifier-free test-time action selection, showing that MAE-guided multiple sampling improves PI-family robustness on LIBERO-Plus with small observed runtime overhead.

Mais de arXiv AI Papers

Entre para seguir esta fonte

arXiv AI Papers·há 3 dias☆

From 1D Benchmarks to Complex 3D Geometries: Can Agents Reliably Specify Problems for Automated PDE Simulation?↗

Turning an engineering request into a simulation requires consistent geometry, physics, boundary conditions, and numerical settings. Recent agent-assisted systems automate simulation setup, but successful execution alone does not establish whether the specification preserves the intended problem or correctly handles missing information. We study whether agents can reliably construct complete executable JSON specifications for automated geometry creation, meshing, PDE solving, and result export. Our workflow combines model extraction with deterministic scientific and capability checks, explicit clarification, and validated local patches that preserve unrelated accepted fields. We evaluate specification correctness, execution completion, and numerical verification separately. On 120 template-related requests, the full workflow achieves 98.3% correct first-turn specification and decision outcomes, compared with 92.5% for schema-assisted single shot and 82.5% for model-only: gains of 5.8 and 15.8 percentage points, respectively. It makes no false-ready decisions among the 42 requests requiring clarification or rejection. Component ablations identify contributions from scientific checks and conditional repair. Separate downstream studies cover 1D, 2D, and 3D steady and unsteady, linear and nonlinear problems, including flow and elasticity with complex three-dimensional wing geometries. A 30-case gallery demonstrates automated execution from retained JSON specifications, while 9 of 12 selected live user-to-results tests achieve joint specification, execution, export, and numerical-reference success (75.0%). These results support traceable, engineer-reviewed simulation assistance within the tested scope, while fresh-response regressions identify consistency as a priority for further improvement.

en

arXiv AI Papers·há 3 dias☆

Rome Wasn't Built in One Turn: Replay-Grounded Reward Signals for Long-Horizon Freeciv Agents↗

Training language-model agents to play long-horizon strategy games is hard because deterministic feedback is sparse and often arrives only at the end of a game. In-game scores can offer dense signals, but they do not reliably predict future outcomes: in 8-68% of position pairs across eight strategy games, the higher-scoring position leads to the worse outcome. We term this phenomenon scoreboard traps. In Freeciv, an open-source turn-based empire-building game where several players manage units, cities, and diplomacy over hundreds of turns, we look into better reward signals for gameplay agents. To facilitate this, we build CivHarness, an open system that saves current game states, branches on different decisions, and replays the game to future turns to provide measurable long-horizon feedback from saved positions. We use CivHarness to train CivTelescope, an LLM Judge that predicts which one of two game positions will yield the better outcome. On scoreboard-trap pairs, CivTelescope is 1.3-2.0 times more accurate than off-the-shelf language models, and it generalizes to new game starts, later turns, and unseen rules. We then train a Freeciv gameplay agent with CivTelescope as a dense reward, and call this recipe CivMarshmallow. When used in the middle stage of the game where scoreboard traps are common, CivMarshmallow reaches an average training phase gain of 8.94 points, 77% more than the best baseline, the end-of-game reward. Over the full 120-turn game, CivMarshmallow uses a hybrid reward that starts from the scoreboard or the end-of-game reward and shifts to CivTelescope as the game advances; this hybrid reward scores highest among the rewards we compare.

en

arXiv AI Papers·há 3 dias☆

MEMIG: Self-evolving Memory Construction for Antibody Optimization↗

Memory has become central to self-evolving LLM agents, which store and retrieve experience from past rollouts to guide future decisions. Existing systems represent this experience as free-form natural language and maintain it by rewriting, which fits tasks whose experience is narratable. However, in scientific discovery and optimization, experience is often a quantity measured by an oracle under a specific context. Rewriting it as text leaves no index for where a record should be, replaces a measurement with a description, and cannot express how interventions interact. To address this, we argue that agent memory for biomolecule optimization should instead be indexable, quantitative, and structured, and we introduce MEMIG. This memory construction framework realizes these three properties. In detail, MEMIG addresses each insight with a domain-grounded identifier and accumulates the quantitative effect measured by a property oracle, so that updates remain local and retrieval becomes an exact lookup. Next, MEMIG connects insights whose interventions are applied together and measures their epistasis by decomposing each evaluated set. Our results show that MEMIG improves performance in most settings across in silico antibody optimization tasks.

en

arXiv AI Papers·há 3 dias☆

A Unified Framework for Probability Estimation of Data-Dependent Events↗

We study the general problem of estimating D((S)), the probability that an unknown distribution D assigns to an event determined by a sample S of n i.i.d.\ draws from D through a known mapping . The estimation algorithm knows and S but does not have access to any additional samples from D. This formulation captures classical missing mass, the data-dependent targets of adaptive data analysis and post-selection inference, and the test error of a data-trained predictor, whose estimation was recently studied by {Gastpar et al.\ (2024a,b)}. Within this framework we examine the two classical approaches to estimation. The error of the empirical plug-in estimator is exactly a generalization gap, so generalization bounds transfer to it directly. The leave-one-out estimator, which recovers the Good--Turing estimator, is accurate for every mapping that is stable under the deletion of one sample point, including new structured examples such as neighborhoods of the sample and pierced sets. The main application of our framework is to error estimation for a predictor that combines a base classifier with memorized exceptions, a structure common in theory and in practice. For it, the test error is exactly the missing mass restricted to the error region of the base classifier, a region selected by the same sample. We formulate the general problem of adaptive missing mass estimation to address such dependence and propose a natural hybrid estimator, the Good--Turing estimator restricted to the selected region. We show that the generalization of the base classifier alone does not guarantee the accuracy of this estimator, and derive two bounds on its error: the first one based on the VC dimension of possible error regions and another based on the mutual information of the selected region with the sample.

en