A co-evolving skill library lifts tool-use accuracy from 27.7% to 32.0% — with fewer tool calls, not more

Notes on SPyCE: Skill-Policy Co-evolution for Multimodal Agents (arXiv:2607.13854) — Ru Zhang, Weijie Qiu · July 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill evolution

SPyCE trains multimodal agents that think with images by distilling every successful multi-step trajectory into a two-tier hierarchical skill library, rather than collapsing it into a scalar reward or dumping it into a static memory store. Execution skills capture local visual operations as condition-action-effect rules (rotate this, crop that, zoom on this region); workflow skills capture higher-level orchestration as a bottleneck description plus a sketch for which tools to call and in what order. A separate large model (Qwen3-VL-235B) mines both tiers out of the policy's own rollouts, merging near-duplicate skills by embedding similarity and pruning execution skills whose tracked success rate drops, while workflow skills stay largely fixed once established. During reinforcement learning, the policy first retrieves a matching workflow skill, decomposes the task conditioned on it, then retrieves execution skills that fit — so what gets retrieved directly shapes the next rollout, and that rollout regenerates the library for the round after.

SPyCE beats a GRPO baseline and an RLOO baseline on both Qwen3-VL-4B and 8B backbones, across eight benchmarks: agentic tool use (TIR-Bench), multimodal math and chart reasoning (MathVerse, MathVision, WeMath, ChartQAPro), visual search (V*, HRBench-4K/8K). On TIR-Bench, accuracy rises to 32.0% from GRPO's 27.7% while average tool calls per task drop to 4.83 from 5.13 — the agent gets more accurate while calling tools less, not more. Ablations isolate where the gain comes from. Removing workflow skills costs 4.0 accuracy points and pushes tool calls up to 5.91; removing execution skills costs 3.3 points. Freezing the library so it stops absorbing new rollouts — killing the co-evolution loop while keeping the same starting skills — still costs 3.1 points. The gains come from the hierarchy and from letting the library keep updating, not from either piece alone.

Key numbers

TIR-Bench accuracy, SPyCE vs. GRPO	32.0% vs 27.7%
Avg. tool calls per task, SPyCE vs. GRPO	4.83 vs 5.13
Accuracy lost without workflow skills	-4.0 pts
Accuracy lost with a frozen (non-evolving) library	-3.1 pts
V* visual search accuracy (8B backbone), SPyCE vs. GRPO	84.3% vs 82.2%

What this means for skill authors

The two-tier split is the part worth stealing even outside this RL setup. Execution skills and workflow skills are optimized and evolved on different timelines in SPyCE — one keeps churning, one stabilizes early — because they answer different questions: what do I do right now, versus how do I sequence this whole task. That maps onto a distinction skill authors already feel informally: a reusable crop-and-OCR routine is not the same kind of object as a reusable verify-then-escalate workflow, and conflating them into one flat skill list, as most memory-augmented agent setups do, throws away information the policy could use. The ablation numbers back this up directly — losing either tier costs a comparable few points (4.0 for workflow, 3.3 for execution), so neither is dead weight.

The more pointed claim is the second one: a skill library that stops updating loses almost as much (3.1 points) as losing an entire tier of skills outright. That is an argument against treating any skill store as a finished artifact. A catalog that only retrieves from a fixed snapshot, however well-curated at launch, is competing against agents that keep rewriting their own toolkit from what actually worked this week. It doesn't make a vetted, dual-tool-scanned catalog the wrong instinct — curation solves a different problem, trust and safety, not fit-to-task — but it does mean the vetting has to sit on top of a live feedback path, not just an initial bar that skills clear once.

Read the scope honestly, though. This is a research RL pipeline: a 235B model doing offline skill distillation and a purpose-built SFT-then-RL curriculum. Evaluation stays confined to the Qwen3-VL family, on vision-centric tasks. None of that is available off the shelf to someone authoring a skill today. What transfers is the recipe underneath it — merge-or-add deduplication by embedding similarity, and pruning by tracked success rate — as a lightweight way to keep any skill collection from accumulating cruft, even without a policy to co-train against.

Skills related to this research

Baoyu Image Gen Baoyu Image Gen creates images from text prompts using OpenAI, Google, or DashScope providers. Configure default settings, choose quality levels and aspect ratios, and optionally use reference images to guide generation.★ 331 Canghe Image Gen Canghe Image Gen creates images from text prompts across multiple AI providers including OpenAI, Google, DashScope, and Canghe. Configure your preferred provider, quality level, and aspect ratio, with support for reference images and batch generation when needed.★ 418

Related notes

References

SPyCE: Skill-Policy Co-evolution for Multimodal Agents — Ru Zhang, Weijie Qiu (arXiv:2607.13854, 2026)
Shao et al., DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models (2024) — source of the GRPO baseline
Ahmadian et al., Back to Basics: Revisiting REINFORCE-Style Optimization for Learning from Human Feedback in LLMs (2024) — source of the RLOO baseline