When the industry is still using "bigger is better" to pile up parameters, a paper just presented at ICLR 2026 as an Oral gives a different answer. "In-the-Flow Agentic System Optimization for Effective Planning and Tool Use," published by a multi-institution research team, proposes a new agent architecture called AgentFlow — using only a 7B-parameter base model, it comprehensively surpasses GPT-4o across 10 benchmarks including search, math, and scientific reasoning, with an average lead of 14.9%.
The core innovation behind this is not a bigger model, but a brand-new modular agent framework and online reinforcement learning training method.
Current mainstream Tool-Augmented LLMs commonly use a single monolithic strategy, mixing chain-of-thought and tool calling within a complete context. This monolithic design works when the task chain is short and the variety of tools is limited, but once faced with long-horizon planning or diverse tool calling, the linear growth of context leads to dramatic efficiency decline and weak generalization.
AgentFlow, on the other hand, decomposes the agent into four modules each handling their specialty: Planner (decomposes tasks), Executor (calls tools and returns results), Verifier (checks correctness of intermediate outputs), Generator (memory management and context construction). The four modules coordinate dynamically through evolutionary memory, each doing its specialty work.
More critical is the training method breakthrough. The team proposes Flow-GRPO — an online reinforcement learning algorithm. Traditional methods train after agents interact with the environment offline, leading to learned policies being out of touch with real running environments. Flow-GRPO's core idea is decomposing the multi-turn optimization problem into a series of manageable single-turn policy updates, broadcasting a verifiable trajectory-level final outcome reward back to each turn, helping the planner complete end-to-end optimization in the flow, while stabilizing the learning process through group-normalized advantage functions.
Experimental data shows 7B AgentFlow achieves an average 14.9% improvement on search tasks, 14.0% on agent tasks, 14.5% on math tasks, 4.1% on science tasks — not only surpassing same-tier open-source models, but also the larger-scale GPT-4o. As model parameter scale or inference turn count increases, AgentFlow's advantage shows positive scaling, hinting at the scalability of this training paradigm.
AgentFlow's success reveals an important trend: agent training paradigms are evolving from single-model offline training to multi-module online collaborative optimization. The previous approach of mixing reasoning and tool calling in a single model, all-the-way-through, may be replaced by more architecturally-aware modular methods. For developers, this means rather than chasing bigger foundation models, focus on training framework innovation — with the same 7B parameters, the right method can bring qualitative leaps.