MoMHa for multi-objective LLM harness optimization
Optimizes LLM harnesses for accuracy, safety, and token cost using a joint-reward proposer.
MoMHa is a system that frames harness design as a search problem over accuracy, behavioral safety, and token cost objectives.
Balancing competing objectives in LLM harness design for more efficient and responsible AI deployment.
Not specified
http://arxiv.org/abs/2609.30967v1
MoMHa is a system for optimizing LLM harnesses. It treats harness design as a search problem across accuracy, safety, and token cost, using an agentic proposer to find the best balance.
It addresses the multi-objective nature of harness design, where a single focus on accuracy can lead to unsafe or inefficient systems. By jointly optimizing for multiple criteria, it aims for better overall harness quality.
When only a single objective like raw accuracy is critical and other factors are irrelevant.
- MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and Tokens ↗arXiv · arXiv cs.AI · Sep 25, 2026 · paraphrased
First seen on Subvolts Sep 25, 2026. Updated Sep 28, 2026. Spotted a problem? Tell us.
- MoMHa: Multi-Objective Optimization of LLM Harnesses over Accuracy, Safety, and TokensAnthropic · arXiv cs.AI · Sep 25, 2026
The Playbook: reusable how-tos, prompts, configs and tips pulled out of the news.
- HeuristicAuxiliary environment signals improve LLM authorization accuracyProviding LLMs with auxiliary environment signals can substantially improve their accuracy in authorization decisions. · 30 min
- Case studyUsing LLMs to build tools for iOS location data analysisResearchers used an LLM to build a tool that decoded iOS location data, overcoming LLM analysis limitations. · 60 min
- Case studyUsing Screwtape to find confirmed failures of AI monitorsDemonstrates how a red-teamer can find confirmed failures of an AI monitor without developer infrastructure access. · 120 min
- ToolScrewtape adversarial transcript editing pipeline for AI monitorsPipeline converts real agent transcripts into attacks that evade action-only monitors. · 120 min
- HeuristicUse AI agents to orchestrate other AI agentsHave one AI agent act as an orchestrator, directing other agents and monitoring their progress. · 5 min
- Troubleshooting fixPrevent AI agents from unintended internet actions during testingKeep AI agents offline during internal evaluations to prevent unintended actions and security risks. · 5 min
Questions people ask
- Which AI tools does this work with?
- Any current model. It was seen with Claude in the sources.
- Where did this come from?
- arXiv (arXiv cs.AI). Subvolts collected, normalized and credited it; the link above opens the original.