Grounded Continuation introduces a runtime verifier that classifies each utterance in an LLM conversation into one of eight epistemic operations and uses a symbolic engine to maintain a dependency map of claims and their supports. The verifier checks whether a new continuation is grounded by walking this map, a linear-time process that requires no additional LLM calls. On benchmarks such as ReviseQA and MemoryAgentBench, the verifier improves single-hop accuracy for several QA models, even enabling a 7B model to outperform GPT‑4o when guided by the verifier.
By Qisong He, Jinwei Hu, Xinmiao Huang, Changshun Wu, Yi Dong, Xiaowei Huang
arXiv:2607. 20531v1 Announce Type: new Abstract: Large language model (LLM) agents are increasingly deployed over Model Context Protocol (MCP) servers, yet the benchmarks used to evaluate them score the final answer or a fixed "ground-truth" list of tools, both of which are fragile once the underlying data is live and stateful.
By Jerzy Kami\'nski, Ilya Galyukshev, Artem Kuznetsov, Sergey Chuprin, Kirill Redko, Aidar Shumbalov, Anna Kalyuzhnaya
arXiv:2606. 25449v1 Announce Type: cross Abstract: A language model's memory can be worse than having no memory at all.
By Alex Kwon
arXiv:2608. 09254v1 Announce Type: new Abstract: LLM analytics agents are evaluated on SQL syntax accuracy, but production failures look different: questions with two valid business definitions, questions the warehouse cannot answer, deprecated columns after a schema change, and queries that execute successfully while returning the wrong business number.
By Morris Lee
The paper "trajectory-judge: What Outcome-Only LLM Judges Miss on Agent Trajectories" examines the limitations of outcome-only evaluation for large language model agents. Using a deterministic tool‑using support‑desk environment with a scripted oracle policy and a fault injector, the authors compare five different judging approaches—programmatic rules, outcome‑only, step‑rubric at two model sizes, and a self‑consistency ensemble—on metrics such as detection, step localisation, fault typing, calibration, and cost across 400 trajectories. The study finds that outcome‑only judges miss many silent faults and generate false positives, while step‑rubric judges achieve higher recall with no false alarms but at greater cost, and that none of the judges read the final reply, allowing fabricated promises to evade detection.
"whyItMatters":"The findings highlight that current production‑default outcome‑only evaluations can overlook critical failures in agent behavior, underscoring the need for more nuanced, step‑level judging methods to ensure reliable LLM agent performance."
By Hadi Mohammadi
A language model's memory can be worse than having no memory at all. Give a model a memory that kept a wrong conclusion but dropped the work behind it, and it emits that stale value as a confident answer; give the same model an empty memory and it abstains.
The paper introduces Iris-mini and Iris-pro, two search agents trained at 35B and 397B parameter scales. They use a novel data pipeline that constructs reverse‑engineered multi‑hop queries from web hyperlinks, filters trajectories, and alternates supervised fine‑tuning with reinforcement learning in a process called SFT‑RL climbing. Evaluations on several benchmarks show that, with inference‑time context management, the agents achieve the best open‑source results in their parameter ranges.
By Ziyuan Liu, Hengqi Liu, Zichuan Wang, Yang Qin, Jiachen Liang, Xu Chu, Shaowei Chen, Yuantao Gu, Mu Chuan
The paper reports a case study of a large language model (LLM) coding agent tasked with building a multi‑component data system from a detailed specification. During a single session the agent introduced five defects, which were categorized by violated constraints and detection methods. The study also evaluates the agent’s retrieval‑filtering strategy on the HotpotQA benchmark, showing that filtering to a graph‑identified entity set yields higher recall than unfiltered search, with a statistically significant gap across all tested budgets.
By Phanindra Reddy Madduru
The study investigates why small language model agents tend to repeat a tool call that just failed. By recording the failed call and its error message in the transcript, the authors measure a negative corrective gain—agents are more likely to repeat the failed action, with a drop of about 1.03 nats per token. The problem is traced to the harness design rather than the model’s understanding of errors, and the authors show that replacing the verbatim call with a runtime-generated description of the failure can reduce this backfiring effect by 76%.
By Esmail Gumaan
APEX-EM is a non‑parametric experience memory that stores full procedural‑episodic traces in a typed Procedural Knowledge Graph and retrieves them via semantic search, structural‑signature matching, and graph traversal. It uses a Plan‑Retrieve‑Generate‑Iterate‑Ingest workflow to produce, quality‑gate, and commit experiences, indexing both successes and failures so the agent learns what to reuse and what to avoid. Evaluations on five benchmarks with a shared GPT‑4o backbone show significant performance gains, such as +7.6 pp on BigCodeBench transfer and +1.4 pp on Lifelong Agent Bench, demonstrating that the memory adds to model capability rather than replacing it.
By Pratyay Banerjee, Masud Moshtaghi, Ankit Chadha
Organizations that cannot send data to a cloud API increasingly ask: how good is Text-to-SQL if the model must run on-premises on open weights, and which popular accuracy "recipes" are worth their compute? We answer with an honest, fully reproducible benchmark on the BIRD development split (n=1534, Execution Accuracy), evaluating three open model families across two generations -- Qwen2.
The paper introduces MERIT, a training‑free agent that uses an online dual‑polarity memory of oracle‑verified corrections and failed attempts to improve LLM agent repair without updating parameters. MERIT retrieves only finalized memories, classifies failures, and employs a hybrid lexical‑dense retriever before the frozen model generates revisions. Experiments with Qwen2.5‑7B‑Instruct show that MERIT raises execution accuracy on the Spider benchmark from 66.34 % to 69.79 % and on BIRD from 47.35 % to 48.44 %, with varying evidence across datasets.
By Khang Nhat Hoang Vo, Tam Minh Chu, Anh Trac Duc Dinh, Thuyen Vinh Ha Bui, Tho Quan