arXiv:2606. 13709v1 Announce Type: cross Abstract: We study controlled post-training refusal suppression in routed MoE and hybrid-MoE foundation models, aiming to increase non-refusal target-response behavior while preserving general capability under a compact intervention footprint.
arXiv:2610.22640v1 Announce Type: new Abstract: Model-editing backdoors aim to induce harmful responses under a trigger while preserving ordinary behavior on untriggered input. Recent model-editing a...
arXiv:2407. 00740v2 Announce Type: replace-cross Abstract: As large language models (LLMs) are widely adopted in real-world applications, it has become critical to ensure LLMs satisfy safety constraints, such as non-toxicity and logical consistency, as well as task- and situation-specific constraints.
arXiv:2511. 05852v4 Announce Type: replace-cross Abstract: Knowledge editing (KE) offers a lightweight alternative to retraining for updating large language models (LLMs).
arXiv:2609.16145v1 Announce Type: new Abstract: We study a practical question: can a small correction module fix errors in a frozen language model's outputs without degrading its base capabilities? W...
The paper presents a prompt-based method for minimal-edit grammatical error correction (GEC) that reduces overcorrection in large language models (LLMs). It introduces taxonomy-based instructions, batch prompting to regularize overcorrection, and LLM-assisted prompt optimization, achieving an $F_{0.5}$ score of 78.32 on BEA-2019 with Gemini 3.1-Pro. This approach narrows the performance gap to fine-tuned models while avoiding their infrastructure demands.
arXiv:2607. 01978v1 Announce Type: new Abstract: Online multimodal knowledge editing requires injecting a continual stream of visual-textual corrections into multimodal large language models (MLLMs) with bounded overhead and minimal disruption to unrelated behaviors.
arXiv:2605. 29668v2 Announce Type: replace Abstract: LLM agents acting in structured environments fail in operational rather than conversational ways, and reliability depends on procedural knowledge of the environment.
The paper introduces RIPPLE, a method for adapting workflow-synthesizing agents through prompt-policy editing without retraining the underlying model. RIPPLE diagnoses failed execution trajectories, maps failures to specific policy segments, and restricts edits to those segments. It then evaluates candidate edits in isolation and replays only those that remain safe after composition, achieving up to a 23.1% improvement in validation success on a synthetic benchmark and positive gains on additional language‑model backbones.
arXiv:2609.15559v1 Announce Type: cross Abstract: Grammatical error correction (GEC) evaluation has traditionally relied on reference or edit overlap, which can penalize valid rewrites that differ fr...
The paper introduces RC‑GRPO‑Editing, a region‑constrained Group Relative Policy Optimization framework for flow‑based image editing. It localizes exploration by decoupling initial noise perturbations to reduce background‑induced reward variance and adds an attention concentration reward to keep cross‑attention focused on the intended editing region. Experiments on CompBench demonstrate consistent gains in instruction adherence within the editing region while better preserving non‑target content.
The paper introduces SafeCut, a method for Source-Free Domain Adaptation that uses Vision‑Language models as external knowledge. SafeCut employs a cut statistic to gauge prediction reliability, enabling a dynamic, reliability‑gated supervision between the source‑pretrained model and the ViL model. This approach selectively amplifies correct mutual corrections while suppressing error propagation, achieving state‑of‑the‑art performance on multiple SFDA benchmarks.
The paper presents a training‑free weight‑editing technique that reduces language model compliance with harmful instructions. By applying forward passes on a small set of prompt pairs, the method modifies model weights so that changing or deleting instructions in the prompt cannot reverse the edit. Experiments on five instruction‑tuned models show significant improvements in honest response rates for lying and substantial reductions in sexist responses, demonstrating the approach’s effectiveness across distinct behavioral domains.