Research map / Large language models / Reasoning and chain of thought

ExComm: Exploration-Stage Communication for Error-Resilient Agentic Test-Time Scaling

NeurIPS 2026 · Large language models › Reasoning and chain of thought · approach: Test-time compute scaling · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ExComm, an exploration-stage communication protocol that audits parallel agents' belief states for cross-agent factual conflicts, resolves them via tool-based verification, delivers soft belief updates, and diversifies redundant plans, improving agentic test-time scaling on AIME and GAIA.

Problem

Error propagation in long-horizon agentic test-time scaling: intermediate factual errors persist in agents' belief states and contaminate later reasoning, and existing TTS methods give limited control over this.

Approach

Periodically check parallel agents' beliefs for mutually exclusive factual claims during execution (not after candidate solutions), resolve conflicts with a decoupled tool-augmented verifier loop, deliver only targeted resolutions to involved agents as appended (soft) belief updates, and redirect redundant plans toward orthogonal strategies.

Key results

• ExComm with Gemini-2.5-Flash-Lite reaches 46.2 average accuracy vs 40.5 for the best baseline (Tree Search), and 62.3 vs 57.3 for Qwen3.5-4B (Independent + SR).
• Error recovery rate averages 38.7% (Gemini) and 61.8% (Qwen3.5-4B), above the best baselines.
• With Qwen3-235B-A22B-Instruct, ExComm averages 63.0 vs 59.5 for the best baseline; with N=8 Gemini agents, it reaches 50.5 vs 44.7.

Surprising findings

• Adding the consistency module alone without soft updates slightly degrades GAIA Level 3 accuracy (9.0 to 8.0).
• Mixture-of-Agents yields near-zero trajectory diversity (Self-BLEU-based score about 4.2), while Multi-Agent Debate scores higher diversity than ExComm on its own metric (53.0 vs 60.5 is lower; ExComm is highest on both diversity metrics).

Limitations

• Effectiveness depends on the reliability of the verifier; incorrect resolutions can mislead agents in extreme cases.
• Method is scoped to exploration-stage error control and does not replace post-hoc aggregation or refinement.
• Higher per-sample latency and cost than independent N=4 scaling (latency 10.45 vs 7.94 min for N=8 Independent).

Experimental setup

Datasets: AIME 2024 AIME 2025 GAIA

Benchmarks: AIME 2024 AIME 2025 GAIA

Metrics: majority-voting accuracy error recovery rate diversity score (100 - Self-BLEU) API cost latency

Baselines: Base Agent Sequential Revision Independent Scaling Independent Scaling + Sequential Revision Tree Search Multi-Agent Debate Mixture-of-Agents

Builds on: Co-Sight ReAct ProcessBench Multi-Agent Debate Mixture-of-Agents

Closest papers by meaning

• LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling (NeurIPS 2026 · similarity 0.89)
• Evaluating Test-Time Scaling of General LLM Agents (NeurIPS 2026 · similarity 0.88)
• ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory (ICLR 2026 · similarity 0.88)
• DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems (ICLR 2026 · similarity 0.88)
• UltraHorizon: Benchmarking LLM-Agent Capabilities in Ultra Long-Horizon Scenarios (ICML 2026 · similarity 0.88)
• AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems? (ICLR 2026 · similarity 0.88)