Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning

This is an automated message from the Librarian Bot. I found the following papers similar to this paper.

The following papers were recommended by the Semantic Scholar API

• MixSD: Mixed Contextual Self-Distillation for Knowledge Injection (2026)
• The Piggyback Hypothesis of Generalization: Explaining and Mitigating Emergent Misalignment (2026)
• DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning (2026)
• Can LLMs Judge Better Than They Generate? Evaluating Task Asymmetry, Mechanistic Interpretability and Transferability for In-Context QA (2026)
• Multi-Hop Knowledge Composition is Bound by Pretraining Exposure (2026)
• Vision-Default, Prior-Override: Causal Mechanisms of Perception-Knowledge Conflict in Vision-Language Models (2026)
• How LoRA Remembers? A Parametric Memory Law for LLM Finetuning (2026)

Please give a thumbs up to this comment if you found it helpful!

If you want recommendations for any Paper on Hugging Face checkout this Space

You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend

Thank you for the interesting work. I have a question about the generality and interpretation of the Knowing–Using Gap.

The experiments mainly use Qwen2.5 and LLaMA-3.x models. Although the paper includes a CoT prompting baseline, eliciting step-by-step outputs from a conventional model is not equivalent to evaluating a modern reasoning-oriented model in its native thinking mode. Such models are specifically post-trained with reasoning SFT and RL to perform extended autoregressive reasoning, rather than merely being prompted to “think step by step.”

As a preliminary observation, I tested several recent strong models in their ordinary non-reasoning modes. When previously unseen supporting facts were introduced through the context at inference time, these models could already perform multi-hop reasoning reliably, even over relatively long chains. I understand that this ICL setting differs from retrieving knowledge injected through fine-tuning. Nevertheless, it suggests that multi-hop composition itself may not be the primary bottleneck for stronger recent models once the relevant facts are available in textual form.

This raises a question about reasoning-oriented models. Since the fine-tuned models in your experiments can answer the individual supporting facts correctly, a reasoning model might first retrieve and explicitly generate those supporting facts or the bridge entity, and then use the generated tokens as context to derive the final answer. Such an autoregressive retrieve-then-reason process could potentially bypass the knowledge–circuit misalignment observed under direct answering.

Have you evaluated Qwen3 or Qwen3.5 in their native thinking modes after applying the same knowledge-injection fine-tuning procedure? Since these model families include relatively small variants, they appear well suited for a controlled comparison with Qwen2.5. In particular, have you examined whether they can recover and verbalize the learned supporting facts before composing them into the final answer?

Such an experiment would help clarify whether the observed gap reflects a general limitation of parameterized knowledge integration, or whether it can be substantially reduced by the native autoregressive reasoning behavior of modern reasoning-oriented models.