A harder test than recall
A preprint reports that a three-stage method called GRIN gave a language model stronger results on harder questions about newly supplied facts, while keeping direct fact recall comparable with other approaches. The harder questions required either combining information from more than one source or drawing an inference rather than repeating a single fact.
The work compares two filtered, wiki-style collections of source material. BLANK contains 776 raw corpora, and COUNTER contains 252. COUNTER also tests whether a prior fact leaks back into an answer after the model has been updated.
Why the method is different
GRIN is built in three stages. Stage 1 builds question-and-answer data for supervised fine-tuning from extracted facts. Stage 2 samples diverse question and golden-answer pairs conditioned on the corpus. Stage 3 applies Golden-GRPO, adding the golden answer as an off-policy trajectory, meaning a supplied answer path that was not generated by the model's current sampling.
The distinctive design choice is what happens to that supplied trajectory. Golden-GRPO replaces the importance-weighted off-policy branch with a direct supervised gradient scaled by the off-policy advantage. In plain language, the known answer contributes directly to the reinforcement-learning update, with its influence adjusted according to how well it performed.
The main evaluation used Qwen3-4B on BLANK and COUNTER and ran on two H200 GPUs. A model identified as gemini-3.5-flash judged answer accuracy. For the COUNTER leakage check, fail@k used five sampled responses per single-fact question.
What happened when the facts were in the model
On BLANK, GRIN's average accuracy was 35.11%. It scored 52.49% on single-fact recall, 21.16% on multi-source questions and 31.69% on inferential questions. The pattern was uneven: direct recall was the strongest result, while questions requiring multiple sources or an inference were harder.
On COUNTER, GRIN reported 43.65% average accuracy. The scores were 54.64% for single-fact questions, 34.37% for multi-source questions and 42.38% for inferential questions. Its fail@5 rate was 24.05%. Because that measure counts a failure when the prior fact appears in any of five sampled answers, the figure means old-belief leakage appeared on roughly one in four items.
Document access provided a useful reference point. Without document context, closed-book averages were 4.33% on BLANK and 2.33% on COUNTER. With the documents in context, the gemini-3.1-flash-lite model reached 94.05% on BLANK and 99.51% on COUNTER. The gap is consistent with the benchmark's intended distinction: having source material available is different from answering with knowledge that has been injected into the model.
A result with clear boundaries
The paper also compared GRIN with stripped-down training variants and longer runs. On BLANK average accuracy, versions using SFT, GRPO and LUFFY scored 14.11%, 1.83% and 20.38%, respectively, compared with 35.11% for full GRIN. In a longer-training check, baselines were extended to 50 total epochs but plateaued beyond 15 epochs below GRIN's 43.65% COUNTER average.
A cross-model COUNTER check gave a more mixed result. GRIN's reported scores were 42.38% on single-fact questions, 11.01% on multi-source questions, 37.95% on inferential questions and 30.45% overall, with fail@k at 27.72%. The check broadens the evidence beyond the main model, but it covers only one benchmark and one additional model.
The boundaries matter. The main evidence comes from Qwen3-4B tests on two benchmark collections, three question types and an LLM-based accuracy judge; the cross-model result is limited to COUNTER. That setup does not establish how the method would behave under repeated knowledge injections, on procedural, code, mathematical or structured material, or when lexical matching is a poor measure of correctness.
The authors interpret the results as evidence that Golden-GRPO helps knowledge reach the model's internal reasoning rather than remain tied to the training format. That is an interpretation of benchmark scores, not causal proof. The work is an arXiv preprint, version 1, dated 26 Aug 2026, so the reported advantage remains a result for further testing across models, repeated updates and broader forms of knowledge.
Paper data and sources
Original title: From Memorization to Absorption: Mixed-Policy RL for Continual Knowledge Injection
Authors: Zhibo Hou, Fan Zhao, Zhiyu An, Wan Du
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text