Low-Margin Branch Points in Deterministic LLM Deliberatio
Summary
- Researchers present a white-box audit evaluating local branch susceptibility in deterministic multi-agent LLM deliberation.
- The study tests a five-agent committee under greedy decoding by injecting norm-relative residual displacements at selected token positions.
- Low-margin positions and margin-reducing gradient directions show significantly higher branch sensitivity compared to control conditions.
- Realized branches are tracked to distinguish between textual persistence, preference shifts, and final-decision changes.
Why it matters
Understanding low-margin branch points in deterministic LLM deliberation helps developers evaluate local failure modes and execution stability in multi-agent workflows.