arXiv cs.CL
·
18 hours ago
Researchers introduced OAT, a method for identifying which steps cause failures in LLM-based agent systems by training exclusively on successful trajectories using neural controlled differential equations. The approach achieved 200-5000 times faster performance than prompting-based baselines while improving F1 scores by 20% on in-domain and 7% on out-of-distribution datasets. This enables efficient debugging of agentic systems without requiring costly step-level error annotations on failure data.
arXiv cs.AI
·
18 hours ago
AgentCheck is an open-source workbench that helps developers test and fix LLM agent failures by reproducing tool failures, testing mitigations, and confirming fixes work before deployment. The system evaluated five agents across 120 scenarios, with the strongest agent passing 105 scenarios and the weakest passing 77, revealing that failures often manifest as silent misuse of incorrect tool outputs rather than crashes. A retry mitigation improved timeout error handling from 30% success to 100% on the weakest agent, though stale-data faults remained problematic at 3-4 out of 10 cases.
arXiv cs.AI
·
18 hours ago
Researchers developed REDDIT, a post-training method that corrects timestamp drift in autoregressive speech recognition systems caused by long non-speech spans, while preventing degradation of transcription quality. The method uses replay-based distribution editing and was tested on 15 ASR systems, improving long-gap alignment from 38.7% to 95.0% mIoU on Whisper-tiny while updating only 1.6% of model parameters. This approach enables better timestamped transcriptions without requiring human annotations or separate alignment tools.
arXiv cs.AI
·
18 hours ago
Claude Code agentic tools compress session histories into summaries that treat partial output from timed-out commands (exit code 143) as confirmed results, propagating false information across sessions without re-verification. The system conflates terminal observations with durable storage, causing the same information to be incorrectly inherited and relied upon in subsequent sessions. Workflows depending on agentic session continuity for data processing or scientific computation face reliability risks from these unverified results being treated as ground truth.
arXiv cs.AI
·
18 hours ago
Researchers propose Deep Interaction, a method allowing users to directly edit and correct errors in reasoning steps of large language models rather than requesting regeneration. The approach achieves a 25% improvement in correction success rate and reduces token usage by 40% on STEM reasoning tasks compared to baseline methods. This enables more efficient refinement of model outputs when reasoning errors occur in multi-step problem-solving.
arXiv cs.AI
·
18 hours ago
Researchers developed Experience Memory Graph (EMG), a framework that helps LLM agents correct their mistakes by converting failed and successful task trajectories into graphs and extracting correction patterns without iterative prompting. EMG achieved higher success rates than reflection-based baselines on ALFWorld and ScienceWorld benchmarks while requiring no test-time trial-and-error loops. The approach enables agents to recover from failures in a single execution by retrieving and applying learned correction patterns across tasks.
