A Single Revision Step Improves Token-Efficient LLM Reasoning

Large language models (LLMs) achieve higher accuracy on challenging reasoning tasks by scaling test-time compute through multiple trajectory sampling. However, standard aggregation methods like majority voting or individual confidence-based filtering face a fundamental “blind spot”: they evaluate each trace in isolation. As problems scale in difficulty, models often generate hallucinated paths that exhibit misleadingly high confidence, causing the true solution to be suppressed by a narrow margin in traditional voting. We ask: can we enable traces to “peer-review” each other to resolve these near-miss errors? We introduce Packet-Conditioned Revision (PACER), a training-free, inference-only framework that enables reasoning traces to revise their conclusions through a structured coordination step. After a preliminary screening of generated traces, PACER constructs a compact consensus packet containing (i) unique candidate answers, (ii) their aggregated confidence scores, and (iii) representative reasoning summaries for each candidate answer. Individual traces then perform a targeted self-review conditioned on this packet, allowing them to identify specific logical junctions where they diverged from the broader consensus and pivot if their original reasoning is found to be flawed. Final predictions are obtained via confidence-weighted voting over these revised trajectories. On challenging competitive math benchmarks such as AIME and BRUMO, PACER matches or exceeds the accuracy of 256-sample majority voting, significantly outperforming raw ensemble baselines by transforming simple consensus into a collaborative logical refinement process.

💡 Research Summary

The paper tackles a well‑known limitation of test‑time scaling methods for large language models (LLMs) on difficult reasoning tasks: standard ensemble techniques such as majority voting (MV) or confidence‑based filtering treat each sampled reasoning trace in isolation. When problems become hard, models often produce “hallucinated” high‑confidence traces that dominate the vote, suppressing the correct answer by a narrow margin.

To address this, the authors introduce Packet‑Conditioned Revision (PACER), a training‑free, inference‑only coordination framework that enables traces to “peer‑review” each other through a single, low‑bandwidth revision step. PACER builds on token‑efficient sampling methods like DeepConf‑Online, which early‑stop traces based on token‑level uncertainty. The workflow consists of three stages:

1. Stable Pool Construction – A fixed number of sampling attempts (N_try) are performed. Each attempt is monitored with a token‑level uncertainty statistic U_t, averaged over a sliding window to obtain a stability score S(τ). After a warm‑up phase, a percentile‑based threshold s is set; any trace whose stability falls below s is stopped early. The remaining completed traces plus the high‑stability warm‑up traces form a “stable pool” T.

2. Consensus Packet Generation – For every trace in T, the final answer a is extracted (e.g., from a \boxed{} token). The system aggregates answers using confidence‑weighted voting (CWV), where each trace contributes its stability S(τ) as weight, yielding a support score V(a) for each answer. The top‑N answers (A_top) are selected, and for each answer a the most stable representative trace τ*_a is identified. The packet for a consists of three elements: (i) the answer value, (ii) its aggregated support V(a), and (iii) a short rationale extracted from τ*_a. This packet is compact, scaling with the number of top answers rather than the number of traces.

3. Conditioned Self‑Review and Re‑Voting – Each original trace is prompted with the consensus packet and asked to reconsider its conclusion. The prompt explicitly asks the model to compare its own reasoning with the peer‑provided rationales and to switch its final answer if the packet suggests a stronger alternative. This constitutes a single revision round; the token cost is limited to the size of the packet and the brief self‑review generation. After revision, CWV is applied again to the revised set of answers, producing the final prediction.

The authors provide a theoretical analysis (Section 5) introducing a “repair‑vs‑damage” condition: if the packet contains the true answer and the representative rationales are sufficiently discriminative, the expected accuracy after revision is guaranteed not to decrease, and typically improves.

Empirically, PACER is evaluated on competitive math benchmarks: AIME (2024/2025), BRUMO (2025), and HMMT (2025). Results show that PACER matches or exceeds the accuracy of a 256‑sample majority vote while using far fewer generated tokens (≈30‑40 % reduction). Notably, on HMMT 2025 PACER improves over the strong DeepConf‑Online baseline by +10 absolute percentage points (28/30 vs. 25/30). The method consistently dominates the accuracy‑token Pareto frontier, outperforming both pure MV (high token cost) and pure early‑stopping (lower token cost but limited accuracy).

In summary, PACER offers a practical, training‑free solution that bridges parallel ensemble scaling and sequential refinement. By injecting a minimal, structured summary of peer evidence and allowing a single, targeted self‑review, it mitigates the “blind spot” of independent trace evaluation, achieves state‑of‑the‑art performance on hard reasoning tasks, and does so with substantially lower inference cost.