Viswonathan Manoranjan

arXiv2026

SHARD: Safe and Helpful Alignment via Self-Reframing Distillation

Large language models often struggle with sensitive prompts. They may refuse outright, provide generic safety boilerplate, or fail to address the user's legitimate informational needs that can be answered safely. We introduce SHARD, a self-reframing distillation method to improve safe-helpfulness. It first rewrites sensitive prompts to surface benign intent using philosophical guidelines, then reframes its original responses into safe, more helpful ones, and finally fine-tunes the model on its self-reframed responses. Across DNA and the English subset of LINGUASAFE, SHARD improves helpfulness for most model families while preserving safety. It also remains competitive with distillation from a larger teacher model, suggesting that models can internalize safe and helpful behavior elicited from their own. Warning: This paper contains content that may be offensive or harmful.

cs.CL↗

arXiv2026

When Identity Overrides Incentives: Representational Choices as Governance Decisions in Multi-Agent LLM Systems

Multi-agent systems built on large language models are increasingly deployed in strategic policy and governance settings, where agents representing stakeholders with conflicting interests must coordinate under shared constraints. These systems typically assign role-based personas to agents, describing their motivations and objectives. Whether agents with role-based identities follow explicit payoffs or their assigned roles in strategic decision-making remains untested. Here we show that assigning role-based personas suppresses payoff-aligned behavior in four-agent strategic games, shifting equilibrium attainment by up to 90 percentage points even when agents have complete payoff information. We test a 2x2 factorial design (persona presence x payoff visibility) across four models (Qwen-7B, Qwen-32B, Llama-8B, Mistral-7B), and 53 environmental policy scenarios with two equilibria: Tragedy of the Commons, where individual payoff dominates, and Green Transition, where collective payoff dominates. With personas present, all models reach near-zero Tragedy equilibrium in the Tragedy-dominant scenarios despite complete payoff information, and 100% of equilibria correspond to Green Transition. No model reaches Tragedy equilibrium by removing personas alone; only Qwen models reach 65-90% Tragedy equilibrium rates when personas are removed, and payoffs are made explicit. Three distinct behavioral profiles emerge: Qwen shifts equilibrium selection based on framing condition, Mistral increases response variance without reaching the Tragedy equilibrium, and Llama holds near-constant across all conditions. Representational choices in multi-agent LLM systems are governance decisions: persona assignment determines which equilibrium a simulation produces, independent of the underlying incentive structure.

cs.MA↗