{
  "schema_version": "1.1",
  "id": "s13:https://arxiv.org/abs/2609.30218v1",
  "slug": "2609-30218v1-0xtyd0y",
  "url": "https://feed7.dev/p/2609-30218v1-0xtyd0y",
  "title": "Minimally Invasive Steering of Language Models",
  "why_included": "MISVO steers frozen models toward test-time rewards while penalizing output-distribution disruption, offering a parameter-free adaptation method tested on preference and code generation.",
  "summary": "MISVO adds position-specific vectors before the output layer and penalizes them using local KL geometry. It works with frozen models of roughly **1B–14B parameters** and does not update their weights.",
  "practical_implication": "Builders operating their own models could treat steering as constrained test-time adaptation: optimize for a task reward while explicitly limiting changes to the token distribution, instead of relying on unconstrained activation edits.",
  "agent_context": "MISVO adds position-specific vectors before the output layer and penalizes them using local KL geometry. It works with frozen models of roughly **1B–14B parameters** and does not update their weights.\n\nBuilders operating their own models could treat steering as constrained test-time adaptation: optimize for a task reward while explicitly limiting changes to the token distribution, instead of relying on unconstrained activation edits.\n\nMISVO had the highest mean reward in **six of seven settings**, while diversity and coherence stayed near Best-of-N. The material covers preference and code-generation tasks but does not establish deployment cost or behavior outside those settings.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2609.30218v1",
    "published_at": "2026-09-24T17:46:46.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [
    "coding"
  ],
  "topics": [
    "reasoning"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "MISVO had the highest mean reward in **six of seven settings**, while diversity and coherence stayed near Best-of-N. The material covers preference and code-generation tasks but does not establish deployment cost or behavior outside those settings."
  ],
  "connected_context": {
    "meaning": "MISVO adds a middle path between choosing a different model and retraining one: optimize frozen-model behavior at inference time while explicitly constraining distributional drift. Its reward gains with near-Best-of-N diversity and coherence make constrained activation steering testable for owned models, but deployment cost and reliability beyond seven preference and coding settings remain open.",
    "corpus_size": 875,
    "generated_at": "2026-09-25T09:07:29.967Z",
    "connections": [
      {
        "title": "Post-Training Language Models for Gold-Medal Performance in Coding Competitions",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2609.02849v1",
        "feed7_url": "https://feed7.dev/p/2609-02849v1-1vwmx41",
        "reason": "Both improve coding outputs at test time, but MISVO uses constrained position-specific steering whereas the competition system relies on generate-evaluate-refine search after specialized post-training."
      },
      {
        "title": "The Base Model Is Dead — Varun Singh, Arcee AI",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=xbPriQWXtWM",
        "feed7_url": "https://feed7.dev/p/the-base-model-is-dead-varun-singh-arcee-ai-02hts76",
        "reason": "The base-model discussion emphasizes changing capabilities through data and training; MISVO offers a contrasting adaptation route that leaves model weights frozen and limits output-distribution change."
      },
      {
        "title": "Prefix Sliding for efficient test-time scaling",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.26070v1",
        "feed7_url": "https://feed7.dev/p/2608-26070v1-1sn2iyc",
        "reason": "Prefix Sliding addresses the memory cost of long inference, while MISVO changes token distributions during inference; deployments combining such test-time methods would need to measure their joint cost and behavior rather than infer efficiency from reward alone."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-09-24T17:46:46.000Z",
  "modified_at": "2026-09-24T17:46:46.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2609-30218v1-0xtyd0y",
    "json": "https://feed7.dev/p/2609-30218v1-0xtyd0y.json",
    "markdown": "https://feed7.dev/p/2609-30218v1-0xtyd0y.md"
  }
}