Declaring each tool's effect, idempotence and read/write resources lets a verifier check agent traces without labels or judges. Training a 4B model on such traces cut deletes-without-asking and redundant calls and used about a third of the tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →.
Key takeaways · AI-distilled
Beyond verifying traces, EffectWorld's effect declarations drive a compiler-style optimizer that rewrites wasteful AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → traces into verified training signal for fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition → small models.
On Qwen3-4B, the authors report LoRAA cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.Full definition → adapters raised task success from 53% to 99% in distribution and from 74% to 83% on an unseen domain, and say the gains carry over to other model families.
On the Berkeley Function-Calling Leaderboard, the adapter matched the base model's live accuracy and raised multi-turn accuracy from 21.9 to 26.6, per the abstract.
The authors found refusal templates in widely used public tool-use datasets teach a small model not to call tools, cutting BFCL non-live accuracy from 88.0 to 31.4. Filtering those templates out restored it to 84.8.
Positive examples did the teaching, the authors report: suppression and preference objectives on verified bad traces added nothing over a plain supervised fine-tuning control, while targeted positives changed exactly the behavior they demonstrated.
Terms in this piece · Glossary
LoRA — A cheap way to fine-tune a model by training a small add-on layer instead of changing all of the model's weights.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
articleSkillEffect: Checked Lowering for Memory-Bounded Agent ToolsYinuo Wang, Yiyu Shi
articleHarness Engineering in LLM Tool Use via Agent-Native Reusable Tool PrimitivesHaibo Jin, Suijin Wang, Xucheng Yu, Haojing Luo, Haohan Wang
articleToolRobustBench: Stage-Wise Perturbation Evaluation and Failure Diagnosis for Tool-Calling AgentsYiShan Zheng, Yuan Wu, Yi Chang