LoMC: Localized Multidirectional Correction for Refusal Suppression in Routed Foundation Models
On the map: Mixture-of-experts → Expert routing strategies
Abstract (arXiv)
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. Existing broad direction-based edits can perturb general-purpose computation, whereas support-only expert edits often lack sufficient capacity to correct heterogeneous refusal representations. To address this limitation, we introduce Localized Multidirectional Correction (LoMC), a support-gated intervention framework that follows a support-then-correction execution order: it first identifies a compact edit support, then aggregates prototype correction directions into layer-wise correction directions, and finally applies rank-one layer-wise correction only within the selected support. By using the edit support as a structural gating constraint, LoMC increases correction capacity without expanding the intervention scope. Experiments on text-only and multimodal safety benchmarks across four routed backbones show that LoMC substantially improves non-refusal target-response behavior while maintaining general capability under a compact intervention footprint.
Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Expert routing strategies, one of 12 directions in Mixture-of-experts. Builds on 17 of its 36 references that are in the map. First posted on arXiv 10 Jun 2026.
Builds on
17 of 36 references.
Closest ancestors · 3
Same direction, newest first
- 2026Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMs9 Feb
- 2026RASA: Routing-Aware Safety Alignment for Mixture-of-Experts Models4 Feb
- 2025Steering MoE LLMs via Expert (De)Activation11 Sep
Other references · 14
- 2023Universal and Transferable Adversarial Attacks on Aligned Language Modelscited by 2,914
- 2023MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contextscited by 1,582
- 2025InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiencycited by 1,419
- 2024HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusalcited by 1,330
- 2024Refusal in Language Models Is Mediated by a Single Directioncited by 960
- 2024DeepSeek-VL2: Mixture-of-Experts Vision-Language Models for Advanced Multimodal Understandingcited by 559
- 2024JailBreakV: A Benchmark for Assessing the Robustness of MultiModal Large Language Models against Jailbreak Attackscited by 208
- 2025The Geometry of Refusal in Large Language Models: Concept Cones and Representational Independencecited by 104
- 2025LLMs Encode Harmfulness and Refusal Separatelycited by 75
- 2024VLSBench: Unveiling Visual Leakage in Multimodal Safetycited by 59
- 2024Surgical, Cheap, and Flexible: Mitigating False Refusal in Language Models via Single Vector Ablationcited by 48
- 2025Refusal Direction is Universal Across Safety-Aligned Languagescited by 37
Built on by
Not cited yet.
Similar, not cited
Close in wording, no citation either way
- 2026Efficient Refusal Ablation in LLM through Optimal Transport
- 2026Universal Refusal Circuits Across LLMs: Cross-Model Transfer via Trajectory Replay and Concept-Basis Reconstruction
- 2026Decomposing Refusal Steering in Mixture-of-Experts Models
- 2026From Refusal Tokens to Refusal Control: Discovering and Steering Category-Specific Refusal Directions
- 2026Beyond Shallow Alignment: How Post-Training Methods Determine Refusal Circuits And Steering Robustness
- 2026Dynamic Adversarial Fine-Tuning Reorganizes Refusal Geometry
- 2026Refusal geometry reflects refusal training: diverse refusal prefixes can raise stable rank and weaken refusal vector ablation attacks
- 2026The Geometry of Refusal: Linear Instability in Safety-Aligned LLMs