{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2608.21278v1",
  "slug": "2608-21278v1-0ozp13x",
  "url": "https://feed7.dev/p/2608-21278v1-0ozp13x",
  "title": "CLEAR: Continuous Latent Adapter Routing for Utility-Preserving LLM Safety Alignment",
  "why_included": "CLEAR conditionally activates a safety adapter instead of applying safety tuning to every prompt, reducing harmful completions while limiting benign-task degradation.",
  "summary": "**CLEAR** uses a hidden-state gate to vary the strength of a safety LoRA while keeping the backbone frozen. On Llama-3-8B-Instruct, HarmBench attack success fell from **32.3% to 0.5%**, with up to **7.1 points** more GSM8K accuracy than global SFT or LoRA.",
  "practical_implication": "For builders serving specialized models, conditional adapter routing is worth evaluating when global safety tuning damages benign tasks. Test the gate, adapter, and untouched backbone as separate failure surfaces across both adversarial and ordinary workloads.",
  "agent_context": "**CLEAR** uses a hidden-state gate to vary the strength of a safety LoRA while keeping the backbone frozen. On Llama-3-8B-Instruct, HarmBench attack success fell from **32.3% to 0.5%**, with up to **7.1 points** more GSM8K accuracy than global SFT or LoRA.\n\nFor builders serving specialized models, conditional adapter routing is worth evaluating when global safety tuning damages benign tasks. Test the gate, adapter, and untouched backbone as separate failure surfaces across both adversarial and ordinary workloads.\n\nThe results come from selected safety and utility benchmarks on **Llama-3-8B-Instruct**. They do not establish production robustness, routing reliability under distribution shift, or behavior on other model families.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2608.21278v1",
    "published_at": "2026-08-21T16:36:10.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [
    "security"
  ],
  "topics": [
    "model-selection",
    "agent-reliability"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The results come from selected safety and utility benchmarks on **Llama-3-8B-Instruct**. They do not establish production robustness, routing reliability under distribution shift, or behavior on other model families."
  ],
  "connected_context": {
    "meaning": "CLEAR narrows safety alignment from a global model rewrite to a conditional internal intervention: benign utility can be preserved by routing a frozen backbone through a safety adapter only as strongly as needed. This complements frozen-model reliability mechanisms but adds a new routing surface whose behavior must be evaluated as part of the deployment configuration; the single-model benchmark does not resolve production robustness or false refusals.",
    "corpus_size": 551,
    "generated_at": "2026-08-24T10:05:01.347Z",
    "connections": [
      {
        "title": "You Only Pass Once: Answering and Abstaining Together in a Single Forward Pass of a Frozen Language Model",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2608.14465v1",
        "feed7_url": "https://feed7.dev/p/2608-14465v1-0tpy8xd",
        "reason": "Both preserve a frozen backbone while applying internal interventions; YOPO’s finding that one intervention can corrupt another probe reinforces the need to test CLEAR’s gate and adapter as interacting failure surfaces."
      },
      {
        "title": "Opaque Epistemic Mediation: How LLM Deployment Configurations Shape the Validation of Pseudo-Science",
        "source_name": "arXiv",
        "source_url": "https://arxiv.org/abs/2607.22513v1",
        "feed7_url": "https://feed7.dev/p/2607-22513v1-0yf6na5",
        "reason": "Its deployment-dependent behavior finding makes CLEAR’s gate, adapter, interface, and serving configuration part of the versioned system rather than treating the backbone identifier as sufficient."
      },
      {
        "title": "Announcements",
        "source_name": "Anthropic",
        "source_url": "https://www.anthropic.com/news/fable-safeguards-jailbreak-framework",
        "feed7_url": "https://feed7.dev/p/fable-safeguards-jailbreak-framework-05x4d7q",
        "reason": "Fable 5’s wider false-positive margin supplies a practical utility cost that conditional safety routing is intended to reduce, while also showing that refusal behavior must be tested on legitimate workloads."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-08-21T16:36:10.000Z",
  "modified_at": "2026-08-21T16:36:10.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2608-21278v1-0ozp13x",
    "json": "https://feed7.dev/p/2608-21278v1-0ozp13x.json",
    "markdown": "https://feed7.dev/p/2608-21278v1-0ozp13x.md"
  }
}