Search papers, labs, and topics across Lattice.
3
0
5
8
This work introduces Harness-Zero, which enables harness distillation through agent-as-harness, and shows that for frontier LLMs using the same evolved harness, agent-as-harness outperforms code-as-harness.
The Behavior-to-Value (B2V) task is proposed, a behavior-to-value measurement model based on Value Verification Tuning, which learns to assess whether behaviors provide sufficient evidence for each value inference, and experiments show that B2V-Verifier outperforms strong LLM baselines, improving multi-label classification by 34\%.
LLM agents become *less* like humans as you crank up the reasoning, unless you explicitly model value activation with a Value Verifier trained on real-world interaction data.
