Explainable AI (XAI) methods often oversimplify model internals, leading to a misalignment between explanations and the model, known as unfaithfulness. Faithfulness is defined as the degree of approximation between an explanation and the model, and measuring it rigorously is important for trustworthiness. Current faithfulness metrics suffer from significant limitations, making accurate quantification of this faithfulness either infeasible or computationally expensive. This position paper outlines five major shortcomings of existing metrics. It proposes four research directions to improve the reliability of faithfulness quantification.
EliteTable is a self-improving program-search algorithm utilizing large language models to guide the exploration of executable control policies. It uses a fixed-capacity table of high-scoring solutions to inform new proposals, edits, and remixes of existing solutions. Performance was assessed across six Gymnasium control tasks, with results showing increased candidate-score medians over generations. Gains were observed in some tasks, such as in CartPole-v1, where the selected winner reached the score ceiling quickly. The method demonstrated the improvement of program search and the transfer of gains to new episodes in certain tasks.
A computational study compared Monte Carlo Tree Search (MCTS) and randomized-restart list scheduling for Directed Acyclic Graph (DAG) workflows under an equal budget. MCTS performed 2.88% worse than the restarts, although the deficit was not due to budget starvation. Restarts showed self-improvement on the separate exploration pool, whereas MCTS gains were minimal. The performance difference was workflow-dependent, with restarts outperforming MCTS on sparse irregular workflows. Using textbook UCT instead of the main MCTS rule reduced the deficit significantly.
EvoSyn is a learning-feedback-driven multi-agent framework designed for agentic data synthesis in closed-loop learning. It addresses the gap where observed weaknesses do not directly inform future synthesis, by organizing synthesis around a revisable state of interaction factors. A multi-agent system translates learner feedback into synthesis revisions, guided by a Designer allocating synthesis directions and a Planner grounding them into executable experience. This process allows for adaptation toward weak regions and the expansion of the synthesis space to incorporate newly discovered factors. Experiments demonstrated that EvoSyn improved task success and generated higher-quality, broader synthetic experience.
SpireBench evaluates the long-horizon strategic play of agentic models using the commercial deck-builder game Slay the Spire. The benchmark tests how models make sequential decisions over extended gameplay, assessing stored knowledge, reasoning, and luck management. Frontier LLMs perform better than heuristic baselines and match average human performance in initial phases, but performance degrades as the run lengthens. The study indicates that performance differences stem from how agents manage risk across the entire run, rather than simple memory retention. SpireBench provides a comprehensive, unsaturated benchmark with tools to pinpoint where and why agents make strategic errors.
Verifiable rewards in reinforcement learning often use binary feedback, which is insufficient for fine-grained debugging quality assessment in software engineering agents. Behavioral Self-Rewarding (BSR) is introduced as a method to allow policies to judge their own trajectories to improve performance. BSR reuses the policy to compare observable debugging behaviors across trajectories from the same task and verifier outcome. It aggregates these pairwise preferences into task-local scores, creating auxiliary rewards that offer contrast without ignoring the initial pass/fail ordering. Experiments on SWE-bench demonstrate that this approach consistently improves performance for Qwen3 models.
The CLEAT framework learns value-guided clarification to assist interactive agents in handling incomplete information. It co-trains a clarification-value controller and a task policy using shared interaction outcomes. Residual query value measures the task utility gained by asking a question instead of proceeding immediately. Paired rollouts from interaction history train the controller to predict clarification value and the policy to select useful questions. This method improves agent performance, achieving significant gains on evaluation benchmarks.
Data-centric scaling of RLVR often results in a "zero-advantage trap" where large datasets offer little useful learning signal. To create effective synthetic data, problems must be calibrated so that some rollouts succeed, some fail, and at least one solution is reachable within the training budget. MagicBeanStalk introduces a recursive data-synthesis algorithm to grow seed problems into new, policy- and budget-calibrated training problems. This method improved performance on math benchmarks and finance tasks by synthesizing problems that require necessary reasoning. The approach successfully generated verifiable problems to escape the limitations of simply pooling large corpora.
The research presents an automated pipeline using Large Language Models (LLMs) to construct hierarchical hardware knowledge graphs from hardware datasheets. This process involves chunk-based property extraction, hierarchical routing, and iterative refinement to handle the fragmented and heterogeneous information in datasheets. The framework utilizes LLMs for extraction and reasoning while ensuring source provenance and traceable execution for reliable results. Evaluation showed the framework achieved 98.8% property faithfulness, significantly outperforming a one-shot baseline of 76.1%. These improvements yield more comprehensive and reliable engineering information for hardware design and verification.
The method aligns two unimodal encoders using only a small number of paired data points through distilled optimal transport. It achieves this by representing all modalities based on their similarity to these anchors, placing them in a common coordinate system. A single entropic optimal transport problem solves the matching between the two point clouds, counteracting hubness in nearest-neighbor matching. The resulting plan is distilled into lightweight encoders, making the correspondence inductive. This approach outperforms existing methods in image-text and molecule-text alignment tasks.