multi-hop-factual-qa-qwen2-5-7b-qwen3-8b-given-deliberately-insufficient CLM-0100For multi-hop factual QA with Qwen2.5-7B/Qwen3-8B given a deliberately insufficient retrieved triple subgraph, training a policy with RL to insert bridging triples that let a separate frozen answerer reach the gold answer raises answer accuracy over standard GRPO and over prompting methods like Self-Ask and RECITE, and the improvement carries over to direct chain-of-thought answering with no triples inserted at test time.
Capability: Stating false facts confidently
Observed on
Instruction-tuned Qwen models 0.6B-14B; multi-hop QA (2Wiki, HotpotQA, MuSiQue) plus held-out single-hop sets; context is a sparse two-hop subgraph with any triple containing the g.
Sources
- Measured on seven benchmarks and two main backbones, with a scaling sweep over six Qwen sizes and stage ablations; the frozen answerer makes the inserted triples the only varying input, and >80% of inserted triples were judged by GPT-4o as not derivable from the subgraph. Weaknesses: absolute accuracies are low (20-45%), margins over the strongest baselines are often 2-3 points with no seeds or intervals, the attribution and scoring both rely on the same GPT-4o judge, and the extra priming stage is not compute-matched against baselines.
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 injecting new or counterfactual facts into a 3-4B instruct model by post-training, supervised fine-tuning on extracted QA pairs binds facts to their training question form — recall of the single fact stays reasonable while multi-source and inferential questions over the same facts collapse — and adding more SFT epochs does not close the gap, whereas mixed-policy RL that supplies the golden answer as an off-policy trajectory when on-policy rollouts fail recovers much of the generalization.Absorbing and flexibly using new knowledge · 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 adapting Qwen3 instruction-tuned models from 0.6B to 32B to tool-calling corpora, LoRA supervised fine-tuning gives the best in-distribution exact-match accuracy (best in 15 of 18 settings, never below zero-shot), while GRPO holds only a sub-one-point average edge on cross-dataset transfer and SFT followed by GRPO under a matched data budget is rarely best at either.Using the tools it is given · 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 factual paired-candidate QA where the model already holds the needed knowledge, hallucination can be detected by removing the query keyword that most drives the chosen answer and seeing whether the prediction systematically recovers to the alternative — erroneous answers show that recovery while correct answers merely become ambiguous — but the signal weakens when the knowledge is absent and largely fails on multi-step arithmetic reasoning, where errors are distributed across steps rather than bound to one keyword.Stating false facts confidently · 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: Under the same retrieval budget and RL compute, a plain GRPO run matching or beating the two-stage priming pipeline on the same benchmarks, or the CoT-only transfer gain vanishing when the compute given to the baseline is matched. Proposed technique, not catalogued: Priming with verifiable bridging triples.