Research map / Large language models / Language, safety and interpretability
Learning to Steer: Input-dependent Steering for Multimodal LLMs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes L2S, a small MLP that predicts input-specific steering vectors for MLLMs from hidden states, trained to reconstruct contrastive-prompt (P2S) steering vectors, to reduce hallucination and enforce safety.
Problem
Existing steering for multimodal LLMs uses a single static steering vector (mean steering) regardless of input, which fails when the desired behavior depends on the input; input-specific contrastive prompts needed for such steering are unavailable at test time.
Approach
Compute input-specific steering vectors from contrastive prompt completions (P2S) offline, then train a lightweight 2-layer MLP to predict these vectors from the input's last-token hidden state, enabling test-time input-dependent steering without generating contrastive prompts.
Key results
- On POPE (LLaVA-v1.5, random subset) L2S raises accuracy from 82.73% to 86.46% and F1 from 90.55% to 92.74%, outperforming Prompt, Norm-Rnd and Mean-S.
- On MMSafetyBench (LLaVA-v1.5), L2S reduces Unsafe-score at p>=0.5 from 0.276 (no steering) to 0.082 and raises ED-score from 0.250 to 0.395.
- On CHAIR (500 COCO images, LLaVA-v1.5), L2S lowers CHAIR_s from 17.31 to 16.10 and CHAIR_i from 52.80 to 51.80, with a 64.20% Gemini win rate over no-steering.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use input-dependent L2S steering instead of no steering | LLaVA-v1.5-7B · MMSafetyBench test | Unsafe-score (p ≥ 0.9): 0.204 -> 0.034 | Helped |
| Use input-dependent L2S steering instead of Mean-S | LLaVA-v1.5-7B · MMSafetyBench test | ED-score: 0.329 -> 0.395 | Helped |
| Use L2S instead of no steering | Qwen2-VL-7B · MMSafetyBench test | ED-score: 0.184 -> 0.592 | Helped |
| Use L2S instead of no steering | LLaVA-v1.5-7B · POPE Random | Accuracy: 82.73 -> 86.46 | Helped |
| Use L2S instead of no steering | LLaVA-v1.5-7B · POPE Popular | Accuracy: 80.40 -> 82.58 | Helped |
| Use L2S instead of no steering | LLaVA-v1.5-7B · POPE Adversarial | Accuracy: 76.82 -> 77.76 | Helped |
| Use L2S instead of no steering | Qwen2-VL-7B · POPE Popular | Accuracy: 89.26 -> 91.13 | Helped |
| Use L2S instead of no steering | LLaVA-v1.5-7B · COCO validation captions | CHAIRs: 17.31 -> 16.10 | Helped |
| Use L2S instead of no steering | LLaVA-v1.5-7B · COCO validation captions | Gemini Win Rate: 35.80% -> 64.20% | Helped |
| Increase steering magnitude α for L2S (ablation) | LLaVA-v1.5-7B safety steering · MMSafetyBench | Unsafe-score (p > 0.7) | Helped |
| Increase steering magnitude α beyond the response-quality threshold (ablation) | LLaVA-v1.5-7B safety steering · MMSafetyBench | Response quality | Hurt |
| Set safety steering magnitude to α=2.2 (ablation) | L2S · MMSafetyBench | Response quality | No clear effect |
Surprising findings
- Mean steering degrades or fails to consistently improve hallucination mitigation on POPE, since averaging contrastive directions across varied hallucinated objects cancels out signal.
- On VLGuard safe queries, L2S leaves 87.4% of responses identical to the default, versus 5.2% for mean steering.
Limitations
- Contrastive prompt pairs are hand-designed and there is no guarantee they are optimal.
- Steering is applied as a single linear shift at one layer, which may limit expressiveness.
- Some failure cases remain: responses can be unrelated to the image or still partially unsafe.
Experimental setup
Datasets: MMSafetyBench POPE COCO val2014 VLGuard
Benchmarks: MMSafetyBench POPE CHAIR VLGuard
Metrics: Unsafe-score ED-score (expert-deferring rate) response quality (Gemini 0-9 rating) accuracy F1 score CHAIR_s CHAIR_i recall Gemini win rate
Baselines: No-steering Prompt Norm-Rnd Mean-S Mean-S(BA) P2S CAST
Builds on: Contrastive Activation Addition (CAA) linear representation hypothesis mean steering
Closest papers by meaning
- Programming Refusal with Conditional Activation Steering (ICLR 2025 · similarity 0.87)
- Focus On This, Not That! Steering LLMs with Adaptive Feature Specification (ICML 2025 · similarity 0.86)
- SteerConf: Steering LLMs for Confidence Elicitation (NeurIPS 2025 · similarity 0.86)
- MIA-Bench: Towards Better Instruction Following Evaluation of Multimodal LLMs (ICLR 2025 · similarity 0.86)
- Personalized Steering of Large Language Models: Versatile Steering Vectors Through Bi-directional Preference Optimization (NeurIPS 2024 · similarity 0.86)
- Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs (NeurIPS 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.