arXiv:2606. 04602v1 Announce Type: new Abstract: As agents grow more capable, legal-domain LLM agents promise to turn document-heavy matters into reviewable work products -- yet reliable deployment faces three obstacles: no large-scale evidence on how today's strongest model-and-harness combinations behave on end-to-end legal matters; no agent architecture adapted to the legal vertical, only general-purpose harnesses; and, in a setting that keeps shifting with new facts, authorities, and deadlines, no mechanism for systems to learn from their own outcomes.
By Hejia Geng, Leo Liu
As agents grow more capable, legal-domain LLM agents promise to turn document-heavy matters into reviewable work products -- yet reliable deployment faces three obstacles: no large-scale evidence on how today's strongest model-and-harness combinations behave on end-to-end legal matters; no agent architecture adapted to the legal vertical, only general-purpose harnesses; and, in a setting that keeps shifting with new facts, authorities, and deadlines, no mechanism for systems to learn from their own outcomes. We address each.
Legal Research Bench (LRB) is a new benchmark comprising 413 open-ended U.S. legal research questions, each paired with a gold answer, supporting authorities, and a binary grading rubric. The study evaluates thirteen advanced language‑model agents using web search, case‑law search, page parsing, and retrieval tools, scoring responses only when all required criteria are met and cited authorities verify. Results show that even the best model, Claude Opus 4.8, achieves full correctness on only 42.9% of questions, with performance varying by legal area and task complexity, and no clear link between more tool calls or inference cost and higher accuracy.
By Katrina Drozdov, Oliver Chen, Langston Nashold, Rayan Krishnan
arXiv:2606. 08932v1 Announce Type: cross Abstract: Rule-following agents tasked with executing policies and regulations often fail via Silent Scope Omission (SSO): a model applies a general rule but silently drops nested exceptions or counter-exceptions, producing outputs that appear compliant yet break on important edge cases.
By Jian Chen, Siyuan Li, Chucheng Wan, Zixuan Yuan
The paper reports that a model can pass fidelity checks—verifying that extracted values match the source—without actually opening a datasheet, due to a hidden constraint that disables tool use. To address this, the authors log every tool call in an agentic benchmark and develop two instruments: a rule‑based failure‑attribution classifier and a silent‑failure detector that flags runs based solely on which tools were invoked. While the detector shows low false positives on clean extractions and recovers all planted faults, its recall against correct tool usage but incorrect answers remains unmeasured, and a partial causal chamber confirms only a subset of claims, highlighting limitations in physical verification.
By Qing Ye, Meng-Hsuan Lin
arXiv:2505. 02763v2 Announce Type: replace-cross Abstract: One of the central promises of legal AI is to automate drudgery -- the formal, repetitive tasks of lawyers' work that consume time without calling for much discretion.
By Matthew Dahl, Eric Mart\'inez
CiteVQA is a benchmark that evaluates multimodal large language models on both answering document questions and providing element‑level bounding‑box citations for the evidence used. It contains 1,897 questions across 711 PDFs in seven domains and two languages, with ground‑truth citations generated via an automated pipeline and strict quality control. The evaluation metric, Strict Attributed Accuracy (SAA), rewards a prediction only when both the answer and the cited region are correct, revealing widespread attribution hallucination in current models.
By Dongsheng Ma, Jiayu Li, Zhengren Wang, Yijie Wang, Jiahao Kong, Weijun Zeng, Jutao Xiao, Jie Yang, Bangrui Xu, Yuhan Wang, Bin Wang, Conghui He
AtomCite is an agentic framework that verifies and corrects page‑level citations in multi‑page documents by parsing answers into claims, checking each claim against the cited page image, and applying a deterministic repair policy. The authors introduce DocCite, the first benchmark for this task, built on MP‑DocVQA and DUDE, containing 928 injected instances and 1,909 verified natural errors. Across Gemini, Claude, and GPT models, AtomCite achieves about 93% verification accuracy and improves citation precision from 34% to 87‑90%, while also enhancing hallucination detection in open‑source models.
By Chen Qian, Yimeng Wang, Yu Chen, Lingfei Wu, Andreas Stathopoulos
arXiv:2509. 00761v4 Announce Type: replace Abstract: Large language models are increasingly deployed for legal question answering, where evaluations typically focus on multiple-choice accuracy.
By Boqin Yuan, Ziqi Wang
The paper examines three open-source agent harnesses—LangChain’s deepagents, Earendil’s pi, and DeepSeek’s dsh—each built from contrasting design philosophies. By analyzing their source code and commit histories, the authors find that the mature harnesses converge on five common architectural elements: a commoditized loop, an append‑only replayable session record, model quirks stored as data, progressive disclosure of context, and explicit extension seams. A fourth harness, used as a held‑out check, also displays all five elements and even reuses another’s implementation, indicating that convergence arises from parallel discovery, diffusion, and literal reuse rather than independent invention. The study notes a missing dimension—external verifiability via a tamper‑evident record—highlighting a future axis for provenance‑sensitive domains.
By Dai Jiahong
arXiv:2607. 20527v1 Announce Type: new Abstract: Agentic LLM systems such as OpenScholar and PaperQA2 read the scientific literature and return cited answers, and both they and their benchmarks already check whether those citations hold, with a fixed attribution model or human graders.
By Taewan Goo, Junsik Kim, Kyulhee Han, GwonYul Jo, Jong-Soo Kim, Tae-Hyung Kim
The paper proposes that self‑evolving large‑language‑model agents can only be safely reviewed if their changes are confined to the runtime harness—instruction text, tool‑call logic, and primitive composition—while the model weights remain fixed. A dual‑loop engine with an admission gate records a hash‑chained diff before deployment, and simulation experiments show that this gate admits only safe changes, preventing error increases and maintaining false‑positive rates. When the gate is removed, many harmful changes slip through and missed flags rise, illustrating the gate’s protective role.
By Ravil Akhtyamov