---
title: "Low-Margin Branch Points in Deterministic LLM Deliberatio"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2611.00546v1"
permalink: "https://feed.astralor.com/items/itm_037d5458d3e4e8c14ec36e6f"
published: "2026-10-31T02:01:39.000Z"
score: 54
channel: "ai"
tags: ["LLM","AI Agent","Low-Margin Branch Points","Deterministic LLM Deliberatio","Deterministic","Because"]
---

# Low-Margin Branch Points in Deterministic LLM Deliberatio

## At a glance

Deterministic multi-agent LLM deliberation can exactly re- play a fixed transcript, yet exact replay alone does not test local resistance to controlled internal interventions. Because each response becomes context for l…

## Key points

- Deterministic multi-agent LLM deliberation can exactly re- play a fixed transcript, yet exact replay alone does not test local resistance to controlled internal interventions.
- Because each response becomes context for later agents, a changed greedy token can redirect a transcript-mediated execution.
- We present a white-box audit of this local failure mode in a five-agent committee under greedy decoding.

## Full text

Deterministic multi-agent LLM deliberation can exactly re- play a fixed transcript, yet exact replay alone does not test local resistance to controlled internal interventions. Because each response becomes context for later agents, a changed greedy token can redirect a transcript-mediated execution. We present a white-box audit of this local failure mode in a five-agent committee under greedy decoding. We prospectively select low- and high-margin token positions, require exact no-op re- play, inject norm-relative residual displacements, and measure first transcript divergence within ten generated tokens. Low- margin positions are more branch-sensitive than high-margin controls, while matched margin-reducing gradient directions branch substantially more often than margin-increasing con- trols. We then continue selected realized branches through the committee to distinguish textual persistence from preference and final-decision changes. The contribution is a controlled audit for local branch susceptibility in this specified protocol, not an estimate of deployment frequency, decision correct- ness, or a phenomenon unique to multi-agent systems.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_037d5458d3e4e8c14ec36e6f