Researchers Advance Legal Case Retrieval and Multimodal Reasoning While Enhancing BM25

Researchers have made significant advancements in various fields, including legal case retrieval, computer-using agents, and multimodal reasoning. A self-evolving framework for rule-driven query rewriting has been proposed, which enhances BM25 without any parameter training. Additionally, a framework for reusable web skills has been introduced, which learns transferable interaction patterns and reduces the average LLM-action count on successful trajectories by 8-10%. Furthermore, a study on the internal lifecycle of code reasoning in LLMs has revealed that models first brew the answer and then diverge into one of four resolution outcomes. A framework for financial multimodal reasoning has also been proposed, which accumulates financially grounded reasoning experience from prior trajectories and distills successful strategies and failure-derived cautionary rules into a persistent memory bank.

A benchmark for evaluating long-horizon webpage generation has been introduced, which contains 490 real-world long webpages for structural fidelity evaluation and 507 goal-oriented interaction tasks over 129 webpages for functional evaluation. A framework for strategic resource reallocation has been proposed, which evaluates LLMs on CEO-level strategic resource reallocation and reveals that all models achieve high structural validity but diverge sharply on strategic calibration. A study on the behavior of LLMs has been conducted, which shows that models can improve substantially after training on tens or hundreds of examples of zero. Additionally, a framework for distributed general-purpose agent networks has been proposed, which enables open peer-to-peer networks in which heterogeneous agents can discover one another, establish trust, and execute open-ended tasks.

A benchmark for evaluating personalized workflows predicted by agents has been introduced, which contains 100 tasks across five domains, with 1,246 reference workflow steps grounded in more than 3,900 sources. A framework for proactive preflection and self-evolving memory for zero-shot object goal navigation has been proposed, which enables continuous test-time improvement. A study on the autoregressive curse in long-horizon logical reasoning has been conducted, which shows that small epistemic perturbations introduced early in generation can propagate irreversibly along the Markov decision process flow, triggering cascading failures that drive the reasoning trajectory toward collapse. A framework for dynamic epistemic entropy orchestrated erasable reinforcement learning has been proposed, which eliminates reliance on external signals and enables the model to precisely excise localized logical defects while reusing historical key-value cache streams.