Research map / Multi-agent systems and games / LLM agents

Model Editing as a Double-Edged Sword: Steering Agent Behavior Toward Beneficence or Harm

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

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.

Frames steering LLM-agent ethical behavior as model editing ("Behavior Editing") and introduces BehaviorBench, a three-tier moral benchmark, to show edits can both promote benevolent behavior and induce harmful behavior.

Problem

Steering the ethical behavior of LLM-based agents efficiently and precisely is hard; RLHF-style alignment is costly and coarse, and model editing could be used to alter agent moral behavior in either direction.

Approach

Reformulate agent behavior/moral-judgment steering as a model editing task (behavior-as-target and judgment-as-target editing) and evaluate it with a multi-tier moral benchmark built from existing ethics datasets, measuring both scenario-specific efficacy and global moral accuracy shifts in benevolent and malicious directions.

Key results

• ROME and FT-M achieve the highest efficacy in both malicious and benevolent behavior editing across datasets, while ICE shows varied performance, with benevolent editing via ICE notably stronger than malicious editing.
• Behavior editing induces sustained shifts in global moral accuracy; parameter-modifying methods (ROME, FT-M) outperform ICE, and newer reasoning models such as o1, o3 and Claude 3.7 resist malicious in-context editing better than Claude 3.5 and GPT-4o.
• Edits on Llama3-8B keep BoolQ, NaturalQuestions, GSM8K and NLI accuracy within about 1 point of pre-edit values (e.g., GSM8K 99.60 pre-edit vs 99.56 for ROME malicious).

Surprising findings

• Benevolent in-context editing is substantially more effective than malicious in-context editing, because aligned agents resist unethical instructions.
• Edits to moral behavior leave general knowledge and reasoning benchmarks nearly unchanged, implying the edits are stealthy and hard to detect.

Limitations

• Study confined to textual environments; no sandboxed or physical settings evaluated.
• Does not investigate mitigation or detection strategies for malicious edits.
• Benchmark relies on sampled subsets of existing datasets (about 100 samples per subset) and GPT-4.1-generated open-ended questions.

Experimental setup

Datasets: Social Chemistry 101 ETHICS Jiminy Cricket MoralChoice BoolQ NaturalQuestions GSM8K NLI

Benchmarks: BehaviorBench BoolQ NaturalQuestions GSM8K NLI

Metrics: Efficacy Score (%) moral accuracy (%) accuracy (%)

Baselines: ROME FT-M ICE MEMIT LoRA GRACE

Builds on: ROME MEMIT GRACE EasyEdit LoRA

Closest papers by meaning

• Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.86)
• Think Twice Before You Act: Enhancing Agent Behavioral Safety with Thought Correction (ICML 2026 · similarity 0.85)
• Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.85)
• Instrumental Choices: Measuring the Propensity of LLM Agents to Pursue Instrumental Behaviors (NeurIPS 2026 · similarity 0.85)
• Subliminal Transfer of Unsafe Behaviors in AI Agent Distillation (NeurIPS 2026 · similarity 0.85)
• Can Large Language Model Agents Simulate Human Trust Behavior? (NeurIPS 2024 · similarity 0.85)