knowledge-updating CAP-0026 · activeAbsorbing and flexibly using new knowledge
Covers whether newly injected or post-training facts can be recalled under paraphrase, combined, and used to override stale beliefs; excludes retrieval-based grounding at inference time.
Tags: general
Proposed from the ingestion pipeline rather than chosen by hand. 2 papers independently pointed at this same competence, arriving under 2 different names (knowledge-injection- generalization, knowledge-updating), which is the signal that it is a real recurring topic and not one author's framing. Two papers describe the same gap between storing and actually using new facts, which long- term-memory does not cover.
What counts as this capability
Scope boundary used when deciding whether a paper is really about this capability, rather than merely mentioning it.
Covers whether newly injected or post-training facts can be recalled under paraphrase, combined, and used to override stale beliefs; excludes retrieval-based grounding at inference time.
Claims
- observationsingle paper↑1↓0On 1,010 counterfactual edits rooted in Chinese cultural entities and aligned with English counterparts, an edit made in one language largely does not propagate to the other: edits succeed in the edited language (about 80% for Chinese at batch size 10, falling with batch size) while the other language keeps answering the old fact, so multilingual models store the same fact in language-specific ways that editing does not bridge.
- mechanismsingle paper↑1↓0When 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.
- mechanismsingle paper↑1↓0When injecting revised medical guideline facts into 4B-14B instruct models by LoRA SFT under a matched token budget, rendering the same update events as extended-matching items — several related vignettes sharing one candidate answer pool — transfers better to a held-out temporally anchored oncology staging benchmark than short-answer, fill-in-the-blank, or per-vignette multiple-select renderings of the identical content.
- mechanismsingle paper↑1↓0Knowledge-editing methods such as ROME and fine-tuning insert a stated fact reliably but do not carry its logical consequences: on multi-hop questions generated from knowledge-graph rules that the edit entails, accuracy is up to 24 points lower than on the directly edited fact, so an edited model can assert a new fact and still answer as if its implications never changed.
- observationsingle paper↑1↓0When 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.
- mechanismsingle paper↑1↓0A single fact edit exerts two coupled pressures on its knowledge neighbourhood, propagation to facts that should change and leakage into facts that should not, and planning the edit's target representations jointly under both constraints, with a gate that rejects risky plans before touching parameters, improved propagation and preservation on RippleEdits by at least 7% across backbones over methods that treat the two separately.
Techniques
- Inject the golden answer as an off-policy trajectory during trainingDuring GRPO-style knowledge-injection RL, add the gold answer to each rollout group and apply it as a direct supervised gradient scaled by its group-relative advantage (no importance ratio or clipping), so learning signal survives when all
- Joint neighbourhood optimization for editstrainingBefore writing an edit into parameters, optimize the target representations for the edited fact and its neighbours together, penalizing both failed propagation and leakage, and reject edit plans whose semantic risk is high.
- Render knowledge updates as shared-pool extended matching ittrainingGroup related update vignettes into one training instance with a shared answer option pool so the model must discriminate across vignettes between current and superseded recommendations.
- Update knowledge by synthetic news mid-training plus abstenttrainingA teacher model turns a new-event knowledge graph into long-form synthetic news articles for next-token mid-training, plus preference pairs that reward chain-of-thought answers and punish outdated values, hallucinations, and unwarranted abs
Sources
Every paper, post or observation behind this capability. Ones cited by a claim show the claim and its stance; ones filed directly are about the capability but have no claim drafted from them yet. Open one to add support or contest it.
- about this capability · no claim drafted yet
- ↑ supports a claim here
- ↑ supports a claim here
- ↑ supports a claim here
- ↑ supports a claim here
- ↑ supports a claim here
- CLM-Bench: Benchmarking and Analyzing Cross-lingual Misalignment of LLMs in Knowledge Editing2026-01-24↑ supports a claim here
Capabilities are a way of carving up the subject, and carvings are arguable. Say so if this one is wrong — especially a proposed one, which a pipeline added because several papers used the same framing, not because anyone decided it was right.