training-small-instruction-tuned-models-reinforcement-learning-plus-factual-supervision CLM-0182When training small instruction-tuned models with reinforcement learning plus factual supervision, routing each atomic fact's verification score only to the tokens that produced it, and down-weighting verifier judgements that do not change when their key evidence is removed, improves factuality benchmark scores over trajectory-level or reasoning- step-level factual rewards.
Capability: Checking claims against evidence
Observed on
3B-parameter instruct models (Qwen2.5-3B-Instruct, Llama-3.2-3B-Instruct), GRPO-style RL on multi-hop QA with Wikipedia evidence snippets, GPT-4o for fact extraction and NLI verifi.
Sources
- Two 3B models, four hallucination benchmarks and four math benchmarks; FARCA beats KnowRL, FSPO, FaithRL and plain GRPO by roughly 1.75-2.21 average points. Ablations remove token provenance and reliability estimation separately, each dropping scores, which isolates the two components. Small models only, single seed reported, and the pipeline depends on GPT-4o extraction so the comparison is not compute-matched.
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 same-family pair of reasoning and instruction-tuned 4B models, learning one interpolation ratio per module (attention block, FFN block, embedding, final norm) with a reward that scores correctness and ranks correct responses by length cuts average generated tokens by about 27% relative to the reasoning model while holding average accuracy, and beats fixed global-ratio merging and CMA-ES search over the same ratios.Cutting to the chase · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- For small instruction-tuned models (3-4B) answering closed-form true/false questions under a stated user opinion, scoring each sampled response by how much more common its answer is than the group itself predicted, and using that score as the GRPO reward, reduces answer flips under user pressure and raises accuracy without any labels, roughly matching label-supervised desycophancy fine-tuning at higher compute cost.Telling the user what they want to hear · reviewed by AI
- 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
- In LLM fact-checking over supplied evidence, high veracity accuracy does not imply the verdict depends on that evidence: when the model's own cited sentences are iteratively removed, off-the-shelf models (GPT-4o-mini, Gemini-2.5-Flash-Lite, Qwen-2.5-32B, Llama-3.1-8B) keep their SUPPORT/REFUTE labels at a non-zero plateau instead of abstaining, and training with paired supervision that requires NOT_ENOUGH_INFO on the evidence-ablated input removes most of that persistence without losing accuracy.Checking claims against evidence · reviewed by AI
Notes
Ingested unreviewed on 2026-09-07 and deliberately inert until a human endorses it: it does not move a technique standing, does not count toward the backtest, and is excluded anywhere a claim would carry weight. Drafted confidence: low. Falsifier as drafted: An ablation or replication where token-level routing and reliability weighting give no gain over step-level factual rewards, or where coarse aggregation matches it once tuned equally. Proposed technique, not yet catalogued: fact-aligned reliability-weighted token credit.