Low-Margin Branch Points in Deterministic LLM Deliberatio

Hajime Shimao, Warut Khern-am-nuaarXiv 2026 · arXiv

TL;DR

This paper presents a white-box audit of multi-agent LLM deliberation, revealing that "low-margin" tokens—where the model is nearly undecided between two choices—act as critical branch points. Small internal perturbations at these positions can flip a single token, triggering a cascade that propagates through the shared transcript and potentially alters the final group decision, even in a deterministic system.

What Problem Does This Paper Address?

Multi-agent LLM systems are increasingly used for complex reasoning, where multiple "agents" exchange messages to reach a consensus. These systems are often designed to be deterministic (using greedy decoding) to ensure reproducibility. However, the authors argue that reproducibility is not the same as local stability.

In a transcript-mediated conversation, every token generated by one agent becomes the context for all subsequent agents. If a tiny internal state change—perhaps due to hardware noise, quantization, or cache approximations—causes a model to pick its second-choice token instead of its first, that "branch" might derail the entire deliberation. The paper seeks to quantify how sensitive these systems are to such "micro-interventions" and whether these small changes actually matter for the final outcome.

Methods

The researchers set up a five-agent committee using Qwen2.5-7B-Instruct to discuss a health-financing policy task. They employed a "white-box" auditing approach:

  1. Target Selection: They identified "low-margin" tokens (where the logit gap between the top-1 and top-2 tokens was small) and "high-margin" controls.
  2. Residual Intervention: They injected a small displacement ($\Delta h$) into the residual stream at Layer 27 of the transformer during a specific generation step. This displacement was relative to the norm of the original hidden state.
  3. Boundary Testing: They used backpropagation to find the gradient direction that most effectively reduces the margin between the top two tokens, testing if moving toward the "decision boundary" predictably causes branching.
  4. Continuation Analysis: When an intervention successfully caused a "branch" (a different token was chosen), they allowed the deliberation to continue for the full 20 rounds to see if the change persisted, was ignored, or altered the final committee decision.

Datasets

The study primarily uses a custom health-financing policy task where a commission must choose between three universal-coverage designs (Options A, B, and C).

For robustness, they also tested:

  • Five additional policy scenarios.
  • A smaller model variant (Qwen2.5-3B-Instruct).
  • "Neutral-label" versions where agents were labeled simply as Agent 1 through Agent 5 to remove semantic role biases.

Key Insights and Ideas

  • Logit Margin as a Vulnerability Map: The "margin" (the confidence gap between the best and second-best token) is a highly reliable predictor of susceptibility. High-margin tokens are incredibly robust to internal noise, while low-margin tokens are the "weak links" where branching occurs.
  • The Geometry of Decisions: Branching isn't just about the size of a perturbation, but its direction. Interventions aligned with the margin-reducing gradient were significantly more likely to cause a branch than random noise.
  • Transcript-Mediated Amplification: In a committee, a single flipped token isn't just a local error; it's "committed" to the shared memory. Later agents treat this altered token as a ground-truth part of the conversation, which can lead to "preference propagation" where the group's internal state diverges from the original path.
  • Deterministic != Robust: Even if a system replays perfectly every time, it may exist on a "knife-edge." A perturbation as small as 0.1% of the residual norm can be enough to flip a low-margin token.

Key Results

  • Branching Rates: Low-margin positions branched in 11.8% of trials (604/5,120), whereas high-margin controls branched in 0% of trials (0/2,560).
  • Dose Response: As shown in Figure 3, the probability of a branch increases sharply with the scale of the perturbation ($\epsilon$) for low-margin tokens, while high-margin tokens remain perfectly stable across the same range.
Figure 3: Dose response. Branch probability rises with per- turbation scale at low-margin positions, while high-margin controls remain stable over the grid.
  • Gradient Sensitivity: Reducing the logit margin via gradient-aligned displacement produced branches in 43.3% of cases, compared to only 2.5% when the displacement increased the margin (Figure 4A).
Figure 1: Layer-wise raw hidden-summary divergence for canonical and prompt-surface variants at matched protocol locations. We use this localization statistic only to rank can- didate windows and choose layer 27 for the subsequent causal audit.
  • Decision Impact: In the continuation study, out of 15 realized low-margin branches, four resulted in a change to the agents' internal preference states, and three ended with a completely different final committee decision.

Limitations and Future Work

The authors acknowledge several constraints:

  • Scope: The study focused on one primary model (Qwen 7B) and one specific deliberation protocol.
  • Decoding: They used greedy decoding; results might differ with sampling-based methods (though sampling is inherently non-deterministic).
  • Real-world Noise: The "stress probes" used are controlled interventions, not direct simulations of specific hardware faults or quantization errors.

Future Work suggested includes comparing single-agent vs. multi-agent "amplification" effects, auditing different layers of the transformer, and measuring the impact of real-world perturbations like KV-cache approximations or low-precision quantization on these low-margin branch points.

Summary updated

Abstract

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.
BibTeX
@misc{shimao2026low,
  author = {Hajime Shimao and Warut Khern-am-nua},
  title = {{Low-Margin Branch Points in Deterministic LLM Deliberatio}},
  year = {2026},
  eprint = {2611.00546},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00546},
}
Figure