Co-Sight: Enhancing LLM-Based Agents via Conflict-Aware Meta-Verification and Trustworthy Reasoning with Structured Facts

TL;DR

Co-Sight tackles the verification bottleneck of long-horizon reasoning in LLM-based agents by reframing verification as conflict-aware auditing (CAMV) and grounding reasoning in a provenance-aware structured facts module (TRSF). The approach uses a closed-loop architecture where CAMV localizes verification to disagreement points among diverse expert views, bounded by the size of the conflict set $|S_c|$, and TRSF maintains verified anchors across four fact types within a three-tier context. The combination yields a scalable, transparent reasoning process with a strong empirical punch: GAIA 84.4%, HLE 35.5%, and Chinese-SimpleQA 93.8%, with ablations confirming synergy between structured grounding and conflict-driven auditing. These results suggest a practical path to trustworthy long-horizon reasoning, offering better cost-benefit trade-offs than purely generative improvements, and enabling more auditable reasoning for complex domains.

Abstract

Long-horizon reasoning in LLM-based agents often fails not from generative weakness but from insufficient verification of intermediate reasoning. Co-Sight addresses this challenge by turning reasoning into a falsifiable and auditable process through two complementary mechanisms: Conflict-Aware Meta-Verification (CAMV) and Trustworthy Reasoning with Structured Facts (TRSF). CAMV reformulates verification as conflict identification and targeted falsification, allocating computation only to disagreement hotspots among expert agents rather than to full reasoning chains. This bounds verification cost to the number of inconsistencies and improves efficiency and reliability. TRSF continuously organizes, validates, and synchronizes evidence across agents through a structured facts module. By maintaining verified, traceable, and auditable knowledge, it ensures that all reasoning is grounded in consistent, source-verified information and supports transparent verification throughout the reasoning process. Together, TRSF and CAMV form a closed verification loop, where TRSF supplies structured facts and CAMV selectively falsifies or reinforces them, yielding transparent and trustworthy reasoning. Empirically, Co-Sight achieves state-of-the-art accuracy on GAIA (84.4%) and Humanity's Last Exam (35.5%), and strong results on Chinese-SimpleQA (93.8%). Ablation studies confirm that the synergy between structured factual grounding and conflict-aware verification drives these improvements. Co-Sight thus offers a scalable paradigm for reliable long-horizon reasoning in LLM-based agents. Code is available at https://github.com/ZTE-AICloud/Co-Sight/tree/cosight2.0_benchmarks.

Figures (4)

• Figure 1: Overview of the Co-Sight Architecture Integrating CAMV and TRSF. The framework forms a closed verification loop between multiple expert agents and a meta-verification agent. Each agent comprises a planner, actor, and a shared facts module that maintains source-verified knowledge. TRSF continuously records, summarizes, and validates evidence across agents, producing auditable factual anchors. The meta-verification agent applies CAMV, which localizes reasoning disagreements through constraint-based pruning, consensus anchoring, and conflict auditing. Only contentious nodes are re-verified, while validated anchors guide integrative synthesis of the final answer. This conflict-driven and fact-grounded interaction yields scalable, transparent, and trustworthy long-horizon reasoning.
• Figure 2: Illustration of TRSF. Co-Sight gradually extracts the underlying tool records into new facts, and uses the facts module to verify inconsistencies and check for hallucinations.
• Figure 3: Performance comparison on the GAIA test benchmark.
• Figure 4: Performance Comparison on the HLE Benchmark.