Self-patching identifies activation locations where relocating representations substantially improves failed generalization cases and design a simple heuristic strategy which recovers 58--75\% of the oracle headroom in generalization failure.
Abstract
Fine-tuning LLMs to inject new knowledge faces a critical challenge: LLMs can quickly memorize new facts, yet fail to use them for downstream reasoning tasks. We formalize this failure as the \textit{\textbf{Knowing--Using Gap}}, characterized by an accuracy gap and a temporal lag between memorization and generalization. To understand this phenomenon, we fine-tune LLMs with unseen knowledge and monitor the spatial permeation dynamics of the knowledge internally using a novel intervention technique called self-patching. Self-patching identifies activation locations where relocating representations substantially improves failed generalization cases. These results are consistent with a knowledge-circuit misalignment hypothesis: memorized representations can exist internally but may not be routed to computation-effective layers. To demonstrate the practicality of this diagnostic finding, we design a simple heuristic strategy which recovers 58--75\% of the oracle headroom in generalization failure. Experiments are done cross-domain for the robustness of this finding.
Continual knowledge injection is essential for keeping large language models up-to-date in a fast-evolving world. Existing methods rely on supervised fine-tuning (SFT), which memorizes injected facts in their training format but fails to generalize across paraphrasing, document combinations, and reasoning. To address this, we propose Golden-GRPO Injection (GRIN), a three-stage self-learning framework for continual knowledge injection. Golden-GRPO is a mixed-policy reinforcement learning algorithm designed specifically for knowledge injection, which injects a golden answer to provide learning signal even when on-policy rollouts fail on novel facts. We further introduce Blank and Counter, two document-level benchmarks targeting novel acquisition and counterfactual overwrite respectively, each evaluating single-fact recall, multi-source retrieval, and inferential reasoning. Our experiments establish a clear empirical claim: mixed-policy reinforcement learning enables knowledge absorption beyond what supervised fine-tuning can achieve. GRIN substantially outperforms SFT and mixed-policy RL baselines on the harder question types while matching them on basic fact recall.
This work investigates memorization in transformer-based language models by analyzing their memorization dynamics during training over multiple epochs and finds that memorization is neither a constant accumulation of sequences nor simply dictated by the recency of exposure to these sequences.
Madhur Panwar, Gail Weiss, Navin Goyal et al.· Neural Information Processin...· 1 citation
Vision-Language-Action (VLA) models excel in robotic manipulation but suffer catastrophic performance drops when canonical instructions are simply paraphrased. Although this brittleness is typically addressed through costly data scaling, our probing reveals that the root cause is architectural rather than a lack of semantic understanding. Specifically, we demonstrate that current VLAs successfully retain the correct task identity internally. The failure actually stems from the joint encoding of dynamic visual observations and text, which introduces systematic feature shifts. Because the downstream action policy is highly vulnerable to these variations, it fails to translate the preserved semantics into correct control commands. To resolve this structural bottleneck, we propose Grounded Semantic Re-binding (GSR), an elegant intervention that bypasses unstable joint routing by explicitly fusing independently extracted task semantics with native visual features to train a completely re-initialized action expert from scratch. This targeted intervention dramatically restores paraphrastic invariance using only canonical demonstrations. On the LIBERO-Para benchmark, GSR improves success rates by up to 44.6 percent. It enables lightweight models to rival massively scaled baselines and pushes state-of-the-art models to a new record PRIDE score of 70.4, outperforming the recently introduced large-scale pretrained model Xiaomi-Robotics-0 in instruction generation capabilities. Building on these insights, we also introduce ParaVLA, a natively decoupled 0.33B-parameter model exhibiting near-perfect robustness to instruction rewording. Ultimately, our work proves that robust semantic grounding can be achieved through elegant structural design, bypassing the inefficient brute-force data scaling paradigm.
Recent work has proposed that reasoning and memorization in language models can be characterized by a single representation direction, including methods that keep this direction fixed during reinforcement learning. We test two assumptions behind this view. First, are reasoning-oriented and factual-recall task groups approximately single-direction separable? Second, does the resulting geometry remain stable after GRPO? Using Qwen3-0.6B and a controlled 400-example dataset, we find that a one-dimensional projection can match a full 1024-dimensional linear probe with AUROC = 1.00 on the studied task groups. However, after GRPO, the corresponding direction is substantially reorganized: mean-direction cosine averages 0.453, probe-direction cosine 0.445, while direct representation drift reaches 0.511 at the final layer. Probe AUROC nevertheless remains 1.00. The evidence therefore supports single-direction decodability for the studied task groups but challenges fixed-direction stability: the information persists while its geometric realization changes.
Machine unlearning offers a promising approach to remove unsafe content from Multimodal Large Language Models (MLLMs), yet ensuring the precision of unlearning remains a persistent challenge. One reason is that current MLLM unlearning evaluation paradigms suffer from a critical blind spot: they assess model utility through benchmarks whose representations are distant from the forget set, failing to capture knowledge holes---severe degradation on benign adjacent inputs. To probe knowledge holes in unlearned MLLMs, we construct a benchmark that captures unintended degradation on benign inputs sharing generic patterns with the forget set, and confirm through controlled experiments that they are a systematic consequence of commonly used approaches. Furthermore, to bridge this gap, we propose Selective Protection with Anchored Regularization, which protects generic patterns via anchored activation filtering while reinforcing them through entity-abstracted enhancement. Our experiments on SafeEraser demonstrate that SPAR recovers over 98% of vanilla response quality compared to below 50% for standard baselines---while achieving 0.00% attack success rate and competitive model utility. These results underscore the necessity of more fine-grained evaluation for trustworthy MLLM unlearning.
Junxian You, Junkai Chen, Yuhao He et al.· 0 citations
It is revealed that scaled encoders successfully absorb substantial representation load while still converging on the correct family of solutions, and this ability to adapt to varying surface forms is what truly matters for usability and generalizability in LLM and agentic applications.
P. Rajput, Abdoul Aziz Bonkoungou, Albérick Euraste Djiré et al.· 0 citations