1-5b-7b-qwen2-5-backbones-trained-rl-multi-hop-retrieval CLM-0338For 1.5B-7B Qwen2.5 backbones trained with RL on multi-hop retrieval QA, giving the policy reversible memory operations (summary, backtrack over a stack) with per-action process rewards raises F1 over trajectory-reward agentic-RAG baselines, and removing the two memory actions costs more than removing the planning action.
Capability: Carrying a chain of inference through to its conclusion
Observed on
Qwen2.5 1.5B/3B/7B trained on filtered HotpotQA and 2WikiMultiHopQA; evaluated on multi-hop, open-domain and agentic QA with a Wikipedia search tool; F1 only..
Sources
- Validated by: F1 on seven QA benchmarks against No-RAG, naive RAG, agentic RAG and RL baselines (ReSearch, Search-R1, AEPO, ARPO, Mem1); average 33.55 vs 30.16 best baseline at 3B; ablations drop average to 27.37 without memory actions, 25.43 without rejection, 19.94 without both process rewards.. Measured across three model sizes, seven benchmarks, with component, reward and leave-one-action-out ablations. [truncated]
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
- When training an LLM with policy-gradient RL on tasks that decompose into many dependent sub-problems, rewarding each trajectory segment by how many reference reasoning points it newly reaches beats a sparse final-answer reward by a margin that widens with task horizon, and still learns on problem sets where the base policy never samples a success and sparse reward gives no gradient at all.Completing tasks that take many steps or many hours · reviewed by AI
- When training 4B-8B Qwen3 agents with group-based RL (GRPO, DAPO, GSPO) on sparse-reward long-horizon tasks, injecting hints into the environment's feedback — procedural action guidance in early episode steps and early training, switching to enriched state observations in later steps and later training — raises success on the unmodified environments by a few points on average, while lowering performance on tasks that require questioning whether the request is answerable at all.Completing tasks that take many steps or many hours · reviewed by AI
- For multi-hop QA over 28k-896k token haystacks with frozen Qwen2.5 backbones (7B-72B), summarizing the document in a parallel hierarchical tree — overlapping windows scanned ~5x, raw segments scored 'critical' passed straight to the final answer, and one summarization channel per decomposed sub-question — scores higher than sequential memory-agent and BM25 retrieval baselines without training, and beats an RL-trained memory agent on the out-of-distribution dataset while losing to it in-distribution.Losing information in long inputs · 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
- When 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).Getting more from several models than from one · 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: An implementation where adding reversible memory actions and action-level process rewards gives no F1 gain over the same RL setup with append-only think/search actions and outcome-only rewards, on the same backbones and benchmarks.