SkillGrad: Optimizing Agent Skills Like Gradient Descent

Forum topic · 小凯 · 2026-05-31

Summary

Researchers at Penn State propose SkillGrad, a framework that treats LLM agent skill packages as optimizable parameters, applying a gradient-descent-like iterative process without fine-tuning the underlying model. SkillGrad uses hierarchical skill files (L1 metadata, L2 always-loaded core, L3 conditional resources), derives loss evidence from failed trajectories plus contrastive success trajectories (tasks that failed initially but succeed later), and accumulates recurring diagnostic patterns via a textual momentum mechanism. A layer-aware patcher converts diagnoses into generalized edits with anchor-based deduplication. On SpreadsheetBench Verified, SkillGrad lifts GPT-5.4 from 55.8% (base skill) to 71.1%, beating Trace2Skill by ~3 points, and GPT-4.1 from 36.7% to 54.2%, a 17-point gain. It also improves out-of-domain performance on WikiTableQuestions. Ablations show removing momentum costs 6.67 points and dropping contrastive-success diagnostics costs 4.17 points. Ten optimization iterations cost roughly $6.4 in API usage. The work reframes prompt engineering as parameter optimization, using the LLM itself to bridge discrete success/failure signals and text edits.

SkillGrad, from a Pennsylvania State University team, treats agent skill packages as optimizable parameters and iterates on them like gradient descent: failed trajectories provide the loss signal, contrastive successes provide a protection signal, and a textual momentum mechanism accumulates recurring diagnostic patterns across iterations. No model fine-tuning is needed—only the skill files change.

The Problem with Skill Files

LLM agent "skills" are lightweight domain adaptation: reusable procedural knowledge stored in structured files (SKILL.md) loaded when relevant tasks appear. But auto-generated skills underperform expert-written ones, third-party skills may be outdated or buggy, and existing evolution methods (EvoSkill, Trace2Skill) rely on heuristic reflection without an explicit optimization framework, objective, or safeguards against over-optimization.

Hierarchical Skill Parameters (L1/L2/L3)

| Layer | Content | Loading | |---|---|---| | L1 Metadata | YAML name/description/activation | Routing only | | L2 SKILL.md | Core principles, workflows, pitfalls | Always loaded, kept compact | | L3 Resources | Detailed examples, edge cases | Conditional, on-demand |

After iteration: L2 grows from 40 lines to ~154 lines (converging); L3 monotonically grows to ~13,000 characters by iteration 10—core knowledge compresses into the always-loaded header while edge knowledge sinks into the on-demand tail.

Trajectory-Level Loss Evidence

Beyond binary success/failure, SkillGrad uses two evidence types:

Failed trajectories (currently failing): corrective diagnosis
Contrastive success trajectories (failed initially, succeed now): protective diagnosis of what changed and should be preserved

Training tasks are sampled from the initial skill's failure pool, making contrastive successes naturally abundant at iteration 1.

Textual Momentum

Analogous to SGD momentum, a dual-state structure accumulates knowledge:

Persistent pattern memory (M_t): recurring mechanisms with anchors, occurrence iterations, best actions, fix history
Current batch overlay (O_t): compact update signals per batch

New patterns appear ~3 per iteration early, dropping to 0–1 by iterations 5–9; active patterns stabilize at 4–5, indicating convergence.

Layer-Aware Patching

The patcher performs pattern-level generalization (one generalized edit instead of per-task fixes), layer-aware routing (L2 vs L3), and anchor-based deduplication (kebab-case slugs like inspect-before-edit). Without momentum, L2 word count balloons 29.7% and L3 shrinks 32%, weakening conditional loading.

Results

LLM-generated skill initialization (SpreadsheetBench / WikiTQ):

| Method | GPT-5.4 SB | GPT-5.4 WikiTQ | GPT-4.1 SB | GPT-4.1 WikiTQ | |---|---|---|---|---| | No Skill | 62.50 | 78.57 | 44.17 | 52.86 | | Base Skill | 55.83 | 77.14 | 36.67 | 48.58 | | Trace2Skill | 65.28 | 79.05 | 37.22 | 60.00 | | EvoSkill | 68.06 | 78.09 | 37.22 | 53.33 | | SkillGrad | 71.11 | 82.38 | 54.17 | 73.65 |

Gains are larger for the weaker GPT-4.1 (+17.5 pp on SpreadsheetBench, +16.95 pp over Trace2Skill), suggesting weaker models depend more on skill guidance. With third-party initial skills, SkillGrad still leads (69.44% vs Trace2Skill's 63.89% on GPT-5.4).

Ablations: removing momentum costs −6.67 pp; failure-only diagnostics cost −4.17 pp—cross-iteration stability matters more than per-iteration signal richness.

Hyperparameters & cost: batch size 4 is the sweet spot (72.5% vs 70.0% at 2 and 70.8% at 6). Iteration 10 is the best checkpoint (72.5%); iteration 13 declines to 70.0%, since text updates lack convergence guarantees. A 10-iteration run costs ~$6.4 ± 0.38 (rising from ~$0.35 to ~$0.85 per iteration as momentum context grows).

The Bigger Picture

SkillGrad's core contribution is a perspective shift: skill files become parameters rather than hand-written documents—paralleling the shift from feature engineering to gradient descent, and from prompt engineering to automated prompt/skill optimization. The LLM acts as a differentiable bridge, converting discrete success/failure signals into textual gradients and edits.

Paper Info

Paper: SkillGrad: Optimizing Agent Skills Like Gradient Descent (arXiv: 2605.27760)
Authors: Hanyu Wang, Yifan Lan, Bochuan Cao, Lu Lin, Jinghui Chen (Pennsylvania State University)
Limitations: evaluated mainly on spreadsheet tasks; text updates are non-monotonic with no formal convergence guarantees