llm-agents-under-fixed-react-harness-answering-questions-over CLM-0587For LLM agents under a fixed ReAct harness answering questions over executable database-backed APIs plus document retrieval, accuracy falls sharply as the required reasoning depth grows from 1 to 5 hops (most models lose over half their accuracy), and the surviving failures concentrate in language-mediated steps — entity disambiguation, cross-source grounding, schema alignment — rather than in emitting syntactically valid tool calls.
Capability: Carrying a chain of inference through to its conclusion
Observed on
8,000+ locally hosted APIs over 62 BIRD-SQL domains with ChromaDB document collections; trajectory-level grading by re-executing predicted calls; one ReAct harness, 11 frontier and.
Sources
- Measured: best model 70.4% on single-hop endpoint tasks versus 50-51% on compositional BI APIs and 26.0% on multi-source multi-hop; stage sieve shows tool-name accuracy ~73-93% on Dashboard APIs while grounded-answer accuracy is 47-70%, and 40-75% of errors on API-terminated chains traced to extraction/grounding. Benchmark questions are LLM-generated with a 60-item human quality check (87-96% rated high quality), and the judge for stages 1-2 is GPT-OSS-120B, which is also one of the evaluated models.
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For LLM agents that must compose and then execute domain tool workflows, training separate adapters for the two roles splits the gains cleanly: supervised fine-tuning of the planner on reference tool sequences mainly improves workflow-order fidelity, while node-level RL on argument-match rewards mainly improves tool-argument grounding and final-answer quality, and combining both beats either on a 9B backbone.Planning a sequence of actions before acting · reviewed by AI
- For LLM agents on long-horizon web penetration testing, holding the backbone model fixed and moving global state out of the context window onto a persistent graph of verified facts and fact-conditioned intent edges raises task success most on the hardest tasks, where success depends on facts obtained dozens of steps earlier (88.2% overall and 75.0% hard, vs 44.1% and 25.0% for the strongest shared-plain-text-memory baseline).Completing tasks that take many steps or many hours · reviewed by AI
- For single-agent procedural tasks where the relevant state fits a fixed schema known in advance, replacing the appended conversation transcript with a structured execution state that the model rewrites each step — discarding the reasoning trace once the state patch is validated — holds accuracy flat as the horizon grows while history-appending runtimes degrade, and the gain comes from the structure rather than the shorter prompt.Completing tasks that take many steps or many hours · reviewed by AI
- On text-to-SQL benchmarks, aggregate accuracy hides a steep difficulty gradient: pre-trained-language-model-based parsers resolved roughly 80% of Spider cases overall but only about 50% of its extra-hard cases, and as natural-language understanding improved the residual errors shifted toward database-side problems — large schemas, ambiguous columns, dirty cell values — rather than language parsing.Writing a reusable locator: regex, SQL, XPath · reviewed by AI
- When an agent builds a whole Python repository from natural-language requirements alone, refining component boundaries with explicit cohesion and coupling thresholds until no split or merge is triggered produces higher functionality coverage and test pass rate than letting the LLM decide restructuring actions for a fixed budget of rounds, because unconstrained LLM-decided actions keep over-decomposing past the point of a coherent modular structure.Generating and editing working code · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: A model family whose accuracy stays flat from 2 to 5 hops on these tasks, or stage-wise error analysis showing most failures at tool-name/argument-name selection rather than at grounding and answer extraction.