injecting-new-or-counterfactual-facts-into-3-4b-instruct-model CLM-0405mechanismsingle paperReviewed by AI
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.
Capability: Absorbing and flexibly using new knowledge
Observed on
Wiki-style entity-grounded documents, single round of injection, Qwen3-4B and Llama3.2-3B; lexical (ROUGE-L/exact-match) rewards; not tested on procedural, code, or math knowledge .
Sources
- Validated by: On two new document benchmarks (Blank, 776 corpora; Counter, 252 counterfactual corpora) judged by an LLM, against SFT baselines (PIT, Self-Tuning, Autonomous Learning), vanilla GRPO, and LUFFY mixed-policy RL; GRIN reached 21.16% multi-source and 31.69% inferential on Blank vs Self-Tuning's 11.26% and 7.10%, similar single-fact recall, and the lowest prior-belief fail@k (24.05%); baselines trained to 50 epochs still plateaued below it.. [truncated]
Either way, bring evidence: a link, or what you saw.
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 fine-tuning a base model on instruction data, removing from the training targets any factual claim the base model cannot consistently recall raises the supported-claim rate of later long-form generations, but the gain comes largely from a more conservative response policy — more refusals and fewer supported claims per answer — not from generating more correct facts at equal coverage.Stating false facts confidently · 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 post-training a small LLM (Qwen2.5 0.5B-3B) to produce weekly asset-return forecasts from news, rewarding rank correlation of a full listwise asset ordering gives a dense learning signal where exact-match GRPO gives almost none (all sampled actions scoring zero on ~35% of steps), and yields higher Spearman correlation and backtest Sharpe than the SFT checkpoint, standard GRPO, and prompted commercial models.Predicting future events · reviewed by AI
- When new knowledge is a set of interconnected fictional future events rather than isolated fact edits, updating 4-8B instruct models by mid-training on teacher-generated long-form articles plus preference tuning (preferring reasoned answers over outdated, hallucinated, or needlessly abstaining ones) beats graph-retrieval and in-context editing baselines on multi-hop and causal questions, while remaining worse than in-context methods on exact numeric single-hop recall unless retrieved facts are also supplied.Absorbing and flexibly using new knowledge · 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 SFT-based injection run at matched compute that reaches the same multi-source and inferential accuracy as Golden-GRPO, or Golden-GRPO failing to beat SFT on generalization on other base models or corpora.