SkillGLoW: Procedural-Family Skill Consolidation for Self-Improving Agents on Long-Horizon Task Streams
Ao Yan, Xin Zhang, Jiawei Du, Joey Tianyi Zhou
cs.AI
2026-09-02
GLoW consolidates local skills into one prior per procedural family. Across 12 runs the frozen priors gain 17.2 hard points over no-skill, 3.6x more compact than a per-task pool.
Agents now write textual skills from their own runs and inject them back at test time. The skills live in one of two places: a single document that keeps getting rewritten, or a flat pool with one entry per task. Both work when tasks share a structure. Both break when each task needs a different solution.
The document collapses into generic discipline the model already knows. The pool grows with the stream, and each entry stays bound to the instance that wrote it. SkillsBench already showed the gap: human-authored skills lift pass rate by 16.2 points, while skills a model writes from the task description alone sit 1.3 points below a no-skill baseline. What transfers is the solving procedure shared by a cluster of related tasks, not the prompt text and not the instance-specific path.
GLoW splits skills into two layers, weaves them at solve time, and separates them again at consolidation.
At test time the library is frozen. Qwen3-Embedding-8B retrieves the nearest family prior from the instruction; cosine similarity below 0.45 fails closed to a base prior that is always injected. The prior stays fixed for the episode. The local skill updates from the current task's feedback.
Four streams: Terminal-Bench-Pro (32), SWE-bench Verified (20), ALFWorld (42), LiveMathematicianBench (53). Three frozen models, DeepSeek-V4-Pro, MiniMax-M3, GPT-5.4-mini, for 12 continual-improvement runs. Hard is each benchmark's all-or-nothing flag. Soft is partial credit; on SWE it is multiplicative, so breaking a passing test zeroes the score.
| Setting | hard vs no-skill | soft |
| Global prior only | +17.2 | +13.0 |
| Global + local regen | +18.0 | +14.6 |
All 12 runs are positive. The largest lift is on SWE, where each task is a separate repository: MiniMax-M3 with regeneration +30.0 / +27.7, DeepSeek-V4-Pro +25.0 / +25.0. Two models gain only +3.8 and +5.7 on the math end, where tasks already share a procedure and family granularity has the least to add.
Against SkillOpt, a published single-document optimizer run under the same protocol, GLoW leads 15 of 21 cells, SkillOpt 4, two ties. SkillOpt's four wins are all ALFWorld, the one set with a shared action space. On Terminal, SWE, and math, GLoW's hard margin averages +8.3, +6.7, and +6.9.
Ablating granularity to the two extremes: compressing the whole library into one document averages +2.0 and loses on 4 of 12 cells; local-only is +10.9 and still trails global-only; an AWM-style flat pool with retrieval is +5.0, against +11.2 from one round of family consolidation. The gate made 26 decisions, 19 accept and 7 reject. The library it kept is at +14.7 by the last round; that round's candidates average +9.6 if admitted unconditionally.
The same library, unmodified, lifts ALFWorld validunseen (60 tasks) from 73.9% to 83.9% across the three models, and SWE hold-out (30 unseen repos) from 40.0% to 45.6% for MiniMax-M3, mean of three trials. The committed library is 3.6x smaller than a per-task pool. On ALFWorld, MiniMax-M3 compresses 42 tasks into 15 priors, 13 of them wholly inside the dataset's own class labels.
A lot of skill work asks how to write a better entry. This paper pins the unit of reuse at the procedural family. If a long-horizon, heterogeneous stream is still being served by one skill.md or an unbounded experience log, the controls here say both organizations fail on that distribution.
What is usable is the organization, not a pretrained library. Priors are plain text, so one model's library could in principle be handed to another; that transfer is untested. This is an organizational change, not a new training algorithm. The base model is frozen. The only handle is what goes into context.
The authors flag two open questions: transfer was shown where task categories recur, not across a genuine domain shift, and cross-model inheritance of a prior library was not tried.
The streams are small: 32, 20, 42, and 53 tasks. Most main-table cells are a single run at temperature 0.7; only the SWE hold-out averages three trials. Clustering weights are hard-coded across all four benchmarks. The commit gate spends extra real executions. A bad family hurts: the largest Terminal family mixed a debugger session, write-ahead-log recovery, and a logic-gate CRC32 build, repaired one task and broke another, the only large family in the study with no net gain.
The 17.2-point mean should not be read as a tightly estimated effect size.