Editorial illustration for Alibaba's Metis agent cuts redundant AI tool calls to 2% and boosts accuracy
Alibaba's Metis agent cuts redundant AI tool calls to 2%...
Alibaba's Metis agent cuts redundant AI tool calls to 2% and boosts accuracy
Alibaba’s Metis agent has slashed needless AI‑tool invocations from almost every request—98 % of calls—to a tidy 2 %, while nudging overall correctness upward. The shift isn’t just a tidy engineering win; it reshapes how the system learns. Early on, the model spends most of its compute budget chasing the right answer, a phase the team describes as “accuracy‑first.” Once the agent’s reasoning stabilizes and it reliably lands on the correct solution, the training pressure eases.
That hand‑off between objectives is why the recent drop in redundant calls matters more than the raw percentages. It signals a point where the model can afford to explore efficiency without sacrificing the quality of its output. The following excerpt from the research paper explains exactly how that transition unfolds, laying out the balance between accuracy and reasoning as the agent matures.
Early in training, when the model still struggles with the task, the optimization is dominated by the accuracy objective, forcing the model to prioritize learning correct reasoning and knowledge. As the model's reasoning capabilities mature and it consistently arrives at the right answers, the efficiency signal smoothly scales up. This mechanism causes the model to first master task resolution, and only then refine its self-reliance by avoiding redundant, costly API calls.
To complement HDPO, the researchers developed a rigorous, multi-stage data curation regime that tackles severe flaws found in existing tool-augmented datasets. Their data curation pipeline covers supervised fine-tuning (SFT) and reinforcement learning (RL) stages.
For the SFT phase, they sourced data from publicly available tool-augmented multimodal trajectories and filtered them to remove low-quality examples containing execution failures or feedback inconsistencies. They also aggressively filtered out any training sample that the base model could solve directly without tools. Finally, using Google's Gemini 3.1 Pro as an automated judge, they filtered the SFT corpus to only keep examples that demonstrated strategic tool use.
For the RL phase, the curation focused on ensuring a stable optimization signal.
Metis' ability to slash unnecessary tool calls from 98 % down to 2 % is a concrete win for efficiency. By pruning the blind invocation pattern that typically drags latency and inflates API costs, the agent also nudges its answer quality upward. The researchers attribute the shift to Hierarchical Decoupled Policy Optimization, which steers early training toward pure accuracy before handing more weight to reasoning as the model matures.
Consequently, the system learns to rely on internal knowledge when it can, reserving external tools for truly ambiguous cases. Still, the report doesn't detail how the method performs on tasks beyond the benchmark used, leaving open the question of broader applicability. Moreover, the long‑term impact on overall system robustness remains uncertain.
In short, Alibaba’s experiment shows that disciplined tool selection can coexist with higher precision, but further evidence will be needed to confirm its general usefulness. Future iterations may need to balance this pruning with the risk of missing rare but critical external data. A notable trade‑off.
Further Reading
Related Reading
LLMs & Generative AI LWiAI Podcast #228: OpenAI unveils GPT-5.2, Runway rolls out first world model
LLMs & Generative AI OpenAI's Codex powers Lovable AI, letting millions create apps from text
LLMs & Generative AI Google releases FunctionGemma, a tiny model for natural-language mobile control
From Archives Qiushi Discovery Engine Enables Autonomous Science on Optical Platform
From Archives