acceptodds
Under review as a conference paper at ICLR 2027

From Fact Overwriting to Causal Editing: Knowledge Evolution via On-Policy Self-Distillation

Abstract

While Knowledge Editing (KE) enables efficient updates, its dominant *Static Fact Overwriting* paradigm treats LLMs as discrete databases, forcibly injecting isolated facts. Clashing with un-evolved legacy priors, this triggers **Epistemic Dissonance**—a pathology where obsolete beliefs compel the model to explicitly negate the injected update. Idealized interventions reveal that this is a systemic paradigm limitation rather than mere algorithmic noise, with a zero-distortion stress test yielding a catastrophic 95.6% self-refutation rate. Given the causally driven nature of real-world knowledge, grounding updates in explicit causal narratives effectively collapses this conflict rate to just 6.6%, underscoring the imperative for **a paradigm shift toward Causal Editing**. To this end, we propose **CODE** (**C**ausal **O**n-policy **D**istillation for **E**diting), a two-stage causal editing framework that effectively integrates such causal logic into model parameters by coupling causal bootstrapping with asymmetric on-policy distillation. Experiments on LLaMA-3.1 and Qwen-2.5 show CODE suppresses self-refutation to 1.8% while securing robust multi-hop accuracy (up to 83.5%), with further probing demonstrating genuine causal internalization over surface memorization—elevating discrete fact overwriting into coherent, globally consistent knowledge evolution. Code is available [here](https://anonymous.4open.science/r/CODE-CD3E/).

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Knowledge Editing, Large Language Models, Causal Editing, On-Policy Distillation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_79wZOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper argues that the dominant knowledge-editing paradigm is fundamentally flawed because LLMs store knowledge as interdependent belief networks rather than key-value tables. Injecting an isolated fact without reconciling the surrounding priors triggers Epistemic Dissonance. The paper introduce a new metric Self-Refutation Rate (SRR) judged by an LLM and the pilot experiments show the failure is a paradigm-level property rather than an optimization artifact. Motivated by this, the authors propose CODE. It first employ an open-book Teacher to produce dissonance-free reasoning and bootstrap the Student via SFT on the Teacher's best-scored trajectories. Experiments on LLaMA-3.1-8B and Qwen-2.5-7B (MQuAKE-CF-v2, MQuAKE-T, KnowEdit) show CODE suppresses SRR to 1.8–11.9% while roughly doubling multi-hop accuracy over the best baselines. Meanwhile, CODE scales to 90 simultaneous edits, preserves general capabilities, and faithfully internalizes the trained rationale.

Critical strengths

The method is well-motivated by the diagnosis, and each component is ablated. Also, the experiments contains multi-hop counterfactual editing, temporal updates, batch scaling to 90 edits, a 14B model, single-hop locality on KnowEdit, general-capability retention, efficiency amortization, and a comparison to inference-time scaffolding, which makes the methods very solid and strong.

Critical weaknesses
  • The methodological contribution is primarily a competent instantiation of an existing recipe. The "privileged-context oracle teacher + SFT cold start + on-policy KL distillation" pattern is the emerging on-policy self-distillation paradigm. The genuine novelty here is the diagnosis and the choice of privileged context. Overall, I think this paper is well-structured and findings is important for the model editing field so I decide to give it weak accept.

  • Table 5 demonstrates that CODE internalizes narratives faithfully , which also means every fabricated detail of a synthetic scaffold (invented agreements, archival discoveries, etc.) becomes a parametric belief the model will assert as fact. The locality evaluations (KnowEdit) test whether other attributes of the edited entity are corrupted, but not whether narrative-invented events leak into unrelated contexts. In model editing setting, this is acceptable, but I think the author should discuss this properly.

Questions for authors
  1. For batch edits, are narratives generated independently per edit, and is there any check for cross-narrative consistency (e.g., two narratives inventing mutually contradictory events)?
1 comment

Thank you for the thorough responses and the new experiments. The provided results strengthen the paper. I decide to maintain my score. The work is solid and the rebuttal is responsive, but the contribution is primarily within the knowledge-editing subfield. I prefer more discussion for the potential of the proposed method.

Reviewer_cFmGOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies self-refutation after knowledge editing and proposes CODE, which combines supervised bootstrapping with on-policy self-distillation from a teacher conditioned on an explanatory narrative. It evaluates factual accuracy, multi-hop reasoning, and self-refutation across MQuAKE and KnowEdit benchmarks. The reported results show substantially improved consistency while retaining competitive editing accuracy and general capabilities.

Critical strengths
  1. The paper identifies a useful failure mode beyond factual recall: an edited model can state a new fact and subsequently reject it. The joint outcome breakdown and human validation of the judge make the consistency evaluation more convincing (Appendices A.6-A.7).

  2. The two-stage method is coherent and well motivated: supervised bootstrapping initializes useful trajectories, while on-policy distillation addresses the student's own generations. Ablations support the importance of both stages and the explanatory scaffold (Table 2).

  3. The empirical evaluation is reasonably broad, covering counterfactual and temporal updates, multiple model sizes, locality, and general capabilities. CODE substantially reduces self-refutation while maintaining strong multi-hop accuracy, including comparisons with matched data augmentation (Table 1).

Critical weaknesses
  1. Evidence for the stronger causal interpretation remains indirect. Force-decode leaves the original parameters unchanged, and rationale alignment may reflect learning the supplied narrative; neither establishes that causal explanations, rather than additional informative supervision, are necessary for the improvement.

  2. Knowledge preservation and repeated updating need stronger validation. KnowEdit locality is only 60.4-71.6% (Table 12), while batch tests cover up to 90 non-overlapping updates; these results do not establish that neighboring facts remain stable under repeated or interacting revisions.

Questions for authors
  1. Does CODE retain its advantage against a control using equally informative, length-matched factual context without an explanatory transition narrative?

  2. How do locality and earlier edit retention change under sequential updates, particularly repeated revisions of the same entity or relation?

Reviewer_TrXbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies a generation-level inconsistency following knowledge editing: a model may initially assert an edited fact, subsequently negate it, and revert to its original answer. The authors call this phenomenon Epistemic Dissonance and measure it using a conditional Self-Refutation Rate (SRR) under direct queries and adversarial contextual probes. A forced-prefix experiment demonstrates strong rejection of unsupported target assertions by unedited models, while explanatory narratives substantially reduce this behavior.

The proposed method, CODE, first uses a frozen, narrative-conditioned model to generate and select reasoning trajectories for LoRA-based supervised initialization. It then samples trajectories from a student without narrative access and minimizes token-level forward KL divergence against the narrative-conditioned teacher along those trajectories. Evaluation primarily covers LLaMA-3.1-8B and Qwen-2.5-7B on counterfactual and temporal MQuAKE tasks, supplemented by a 14B model, KnowEdit, batches of up to 90 edits, capability-retention tests, and scaffold perturbations. CODE substantially reduces self-refutation while improving multi-hop performance; augmented configurations achieve a minimum SRR of 1.8% and a maximum multi-hop accuracy of 83.5% in different settings.

Critical strengths

The paper targets an important generation-level failure beyond factual recall and demonstrates meaningful improvements in stability. Explicitly asserting and then rejecting an edited fact undermines the reliability of edited models and is poorly captured by short-answer matching.

The empirical validation addresses several relevant alternative explanations and suggests reasonably robust benefits. Beyond the two primary model families, the paper includes a 14B model, temporal updates, KnowEdit, batch editing, capability retention, scaffold compression and substitution, and human validation of automated judgments. Collectively, these results support the practical value of retaining updated information without inference-time narratives while reducing self-refutation.

Critical weaknesses

Insufficient comparison with directly relevant published work leaves the methodological contribution unclear. Siyuan Qi et al., In-Context Editing: Learning Knowledge from Self-Induced Distributions (ICLR 2025)and Propagating Knowledge Updates to LMs Through Distillation(NeurIPS 2023) already investigate contextual knowledge internalization and propagation through distillation, yet neither is discussed or evaluated. Most baselines lack equivalent narrative supervision, while removing the scaffold also reduces supervisory information. Consequently, it remains unclear whether the gains arise from CODE’s distinctive design or richer supervision combined with established context-distillation mechanisms, weakening the claimed paradigm shift.

The experiments do not establish inherent limitations of static editing or genuine causal internalization. Force-decode inserts an answer prefix into an unedited model; subsequent correction does not demonstrate an inherent flaw in parameter editing. Rationale alignment and alternative-narrative reconstruction establish retention of supplied explanations but do not exclude semantic memorization. Matched non-causal supervision controls and unseen tests requiring the learned mechanism are missing, leaving the central mechanistic interpretation unsubstantiated.

Questions for authors

Please state what CODE adds beyond ICE (ICLR 2025) and the NeurIPS 2023 context-distillation approach.

Does performance deteriorate when only the transition explanation is removed while preserving entity background, old/new-answer statements, and comparable length?

Clarify the sample sizes, the separation of training/augmentation questions from evaluation questions, and whether Table 7 compares non-augmented CODE with augmented CaKE.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.