rl-training-central-agent-multi-agent-harness-1-5b-3b-qwen2-5-backbones CLM-0631When RL-training a central-agent multi-agent harness with 1.5B-3B Qwen2.5 backbones on multi-hop QA, optimizing only the coordinating central policy and leaving sub-agents frozen scores higher than jointly training central and sub-agent policies (average F1 47.79 vs 45.36, accuracy 42.88 vs 40.49 at 3B).
Capability: Getting more from several models than from one
Observed on
Central-coordination multi-hop QA and agentic retrieval harnesses with frozen retrieval/conclusion sub-agents; small (1.5B-3B) open backbones; only one harness topology tested..
Sources
- Validated by: Token-level F1 and accuracy across seven multi-hop QA benchmarks with Qwen2.5-3B, central-only vs all-agent joint training under the same optimization scheme.. Measured: one ablation table, Qwen2.5-3B only, seven benchmarks, two metrics; central-only wins on most but joint training wins on a few individual benchmarks (2Wiki accuracy, MuSiQue F1). No compute-matching discussion and no replication at 1.5B.
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 a small agent policy (Qwen3.5-4B) repeatedly deployed on the same task family with environment-level feedback, distilling only the trajectories whose textual-state guidance is newly or persistently useful relative to state-free execution raises the state-free policy's accuracy more than distilling all state-conditioned trajectories, and the resulting policy keeps improving over alternating rounds of state editing and consolidation.Improving its own scaffolding from failures · reviewed by AI
- When fine-tuning a 14B model on knowledge-graph-derived multiple-choice QA for a single medical domain, attaching supporting triples from the same source text chunk to each target triple during supervision, and repairing residual one-hop failures before reinforcement learning, gives higher accuracy on held-out 3-, 4- and 5-hop questions than training on isolated triples or running RL from an unrepaired checkpoint.Carrying a chain of inference through to its conclusion · reviewed by AI
- For long-horizon repository-level coding agents, a small (4B) critic trained to diagnose each just-completed step with a severity label, error type and causal rationale, and to redirect the next action when severity is critical, repairs more initially-failing tasks with a single branch (14.1% fail-to-pass) than five full resamples of the base agent (13.7%), at about one-fifth the base-agent output tokens.Judging another's work as well as a careful reviewer would · reviewed by AI
- For real-world online RL post-training of large pretrained vision-language-action policies on high-precision manipulation, optimizing the action head toward a paired advantage margin against a frozen reference action, rather than maximizing absolute Q values, yields higher autonomous success on long multi-step tasks, because absolute-Q maximization amplifies early optimistic value errors into policy drift that behavior-cloning regularization alone does not stop.Turning plans into physical action · reviewed by AI
- 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
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: low. Falsifier as drafted: Joint training of central and sub-agent policies matching or beating central-only training on the same benchmarks under matched compute, or the ordering reversing at larger scale or on other harness topologies.