Research map / Large language models / Agents, code and math

Tools as Continuous Flow for Evolving Agentic Reasoning

NeurIPS 2026 · Large language models › Agents, code and math · approach: Long-horizon LLM agents · 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 FlowAgent, which generates a complete future tool trajectory per phase via conditional flow matching in a semantic embedding space, decodes it to discrete tools, and replans after each real observation, evaluated on a plan-level protocol built from tau^2-bench, ToolBench, API-Bank and GTA.

Problem

Step-wise autoregressive tool selection lacks a global view of the remaining plan, accumulates errors over long horizons, and generalizes poorly to unseen tools.

Approach

Reconceptualizes tool chaining as continuous trajectory generation: one conditional-flow ODE solve yields one latent anchor per future plan position, each decoded to a tool by distance-based softmax, with only the first tool executed and the plan resampled after each observation (receding-horizon replanning).

Key results

• On Qwen2.5-7B-Instruct, FlowAgent raises Overall Success from 0.6571 (Standard SFT) to 0.7643 and Tool EM to 0.7745.
• On Llama-3.2-3B-Instruct, FlowAgent reaches 0.7571 Overall Success versus 0.6214 for Standard SFT.
• Training time drops from 48.75 min (ToolRL) to 12.29 min, with average peak GPU memory reduced from 68.36G to 55.14G.

Surprising findings

• Heuristic State Key Majority reaches 0.7643 Tool EM yet only 0.6364 Pred Exec, suggesting tool-name recovery does not imply executable completion.
• Oracle tool selection on retail database operations reaches only 0.7059 success, so argument grounding and state-conditioned operations remain bottlenecks.

Limitations

• Theoretical results are finite-horizon sufficient conditions under assumed Lipschitz, margin, and contraction conditions that are not empirically verified.
• Executable test set is small (140 trajectories) with telecom domain saturated; retail split drives discrimination.
• Retail success drops when replanning frequency decreases, and database operations remain hard to ground.

Experimental setup

Datasets: tau^2-bench ToolBench API-Bank GTA

Benchmarks: FlowPlan plan-level benchmark (tau^2-bench telecom and retail splits)

Metrics: Overall Success Retail Success Tool EM Retail Tool EM Pred Exec Arg Value EM

Baselines: Direct-LLM Standard SFT RAG Planning ToolRL ToolPlanner Agent-R1 AutoTIR

Builds on: conditional flow matching ToolRL GRPO tau^2-bench

Closest papers by meaning

• AutoTool: Dynamic Tool Selection and Integration for Agentic Reasoning (ICML 2026 · similarity 0.90)
• Real-Time Reasoning Agents in Evolving Environments (ICLR 2026 · similarity 0.90)
• DyFlow: Dynamic Workflow Framework for Agentic Reasoning (NeurIPS 2025 · similarity 0.90)
• A$^2$FM: An Adaptive Agent Foundation Model for Tool-Aware Hybrid Reasoning (ICLR 2026 · similarity 0.89)
• HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution (AAAI 2026 · similarity 0.89)
• SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents (NeurIPS 2025 · similarity 0.88)