Research map / Large language models / Instruction and fine-tuning
Unintended Misalignment from Agentic Fine-Tuning: Risks and Mitigation
Dongyoon Hahm, Taywon Min, Woogyeol Jin, Kimin Lee
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune LLM on benign web navigation agent data | Llama-3.1-8B-Instruct · WebDojo | attack success rate (%): 32.88% -> 64.38% | Hurt |
| fine-tune LLM on benign web navigation agent data | Llama-3.1-8B-Instruct · WebArena-Lite | success rate (%): 2.42% -> 22.42% | Improved |
| fine-tune LLM on benign code agent data | Gemini-2.0-flash · RedCode-Exec | refusal rate (%): 19.86% -> 3.15% | Hurt |
| fine-tune LLM on benign code agent data | GPT-4o-mini · RedCode-Exec | refusal rate (%): 40.05% -> 37.01% | Hurt |
| prepend fixed 'I can't' prefix to responses (ablation) | fine-tuned Llama-3.1-8B-Instruct agent · WebDojo | refusal rate (%): 0.0% -> 100% | Improved |
| prepend fixed 'I can't' prefix to responses (ablation) | fine-tuned Llama-3.1-8B-Instruct agent · WebArena-Lite | success rate (%): 22.4% -> 10.3% | Hurt |
| add PING optimized prefix to fine-tuned agent | fine-tuned LLM agents · WebDojo and RedCode-Exec | refusal rate: +66.2% web, +44.6% code on average | Improved |
| add PING optimized prefix to fine-tuned agent | fine-tuned LLM agents · WebArena-Lite and MINT-ALFWorld | success rate: 1.8% degradation for both domains | No clear effect |
| use PING instead of PTST or few-shot prompting | fine-tuned agents · WebDojo and RedCode-Exec | refusal rate | Improved |
| combine PING with WildGuard | PING · RedCode-Exec (code agent) | refusal rate (%): 69.5% -> 80.8% (Gemini-2.0-flash) | Improved |
| combine PING with LlamaGuard3 | PING · RedCode-Exec (code agent) | refusal rate (%): 35.6% -> 38.8% (Llama-3.1-8B) | Improved |
| inject optimized string as user-prompt suffix instead of response prefix (ablation) | PING · WebDojo | refusal rate (%): 79.4% -> 14.3% | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- WildGuard — WildGuard: Open One-stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs (NeurIPS 2024) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Do as I do (Safely): Mitigating Task-Specific Fine-tuning Risks in Large Language Models (ICLR 2025 · similarity 0.88)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (ICLR 2026 · similarity 0.88)
- MLLMs Fail to Refuse when Using Tools Agentically (NeurIPS 2026 · similarity 0.87)
- Just Enough Shifts: Mitigating Over-Refusal in Aligned Language Models with Targeted Representation Fine-Tuning (ICML 2025 · similarity 0.87)
- DRIFT-BENCH: Diagnosing CoopeRative Breakdowns in LLM Agents under Input Faults via Multi-Turn Interaction (ICML 2026 · similarity 0.87)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.87)
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.
