SkillGLoW Limits Skill Reuse to Procedural Families on Long-Horizon Streams

The granularity problem SkillGLoW inherits from earlier skill libraries

Earlier work established that agent skills are an external procedural memory layer loaded at inference time, and that the hard part is not writing skills but organizing them [2]. SkillZip framed this as a reuse-granularity mismatch: exposing a whole skill package is too coarse, while text summarization risks hiding verifiers or dropping input contracts, so it moved to section-level procedural subgraphs with explicit contracts [2]. SkillGLoW attacks the same mismatch at a different level, arguing that the reusable unit is neither the whole document nor the individual task entry but the procedure shared by a family of tasks [1]. Its diagnosis is that a single global document assumes one dominant procedure exists, while a per-task library assumes old entries can be reused wholesale, and on heterogeneous long-horizon streams both assumptions fail simultaneously [1].

The empirical basis for that diagnosis is a set of controls run under the same protocol. On LMB's 15-task validation split, a flat library storing all local skills reached 0.333 soft score with 53 entries, while a family library reached 0.467 with 22 entries, matching a flat library 2.4 times its size [1]. The paper also reports that a merged document yields generic discipline the model has already internalized, with one debugging family prior explicitly noting that its two tasks share only generic engineering discipline [1]. This is an interpretation of why the document form fails, supported by the compressor's own merge notes rather than by an independent annotation.

How procedural-family consolidation works, and what the commit gate actually guards

SkillGLoW encapsulates each task's local experience as a skill card, clusters cards by procedure rather than subject matter using four textual views with the abstract signature weighted most heavily, and compresses each family into one candidate prior retaining only applicability conditions, core solving procedure, and common failure modes [1]. Instance details are filtered out and regenerated per task by a Localize module that depends only on the current task's own execution feedback and is never written back to the library [1]. Admission is decided by a verifier-grounded gate: a candidate revision is committed only when its measured execution value on the task set is at least the anchor minus epsilon, where the anchor is the larger of the standing library's value and the no-skill baseline, floored at the previous round's best measurement [1].

The gate is the mechanism that separates this from prompt-optimization approaches that admit an edit on held-out performance within a single domain [1]. It is also the constraint that limits the claim: admission depends on real execution with a verifier, so the method assumes a harness that can score tasks, and the paper does not demonstrate open-domain deployment without one [1]. The three candidate routes sent into the same gate include append-only repair, whose net effect flips sign across models, up 3.1 for DeepSeek, down 3.4 for GPT, and up 4.8 for MiniMax-M3, which is why the authors adjudicate it by measurement rather than by a fixed priority order [1].

Where SkillGLoW sits against Recuris, SCAFFOLD, and SkillZip

Recuris is the closest precursor in framing: it treats the memory-control layer rather than the weights as the trainable surface of a frozen agent, maintains a verified working state that grounds skill invocation, and emits a structured trace linking states, invoked skills, actions, and outcomes [3]. Its fault-localization experiment shows why structure matters, with macro accuracy rising from 13.0 percent given only the outcome to 64.8 percent given the structured trace, and invocation faults named zero times without the trace against 38.9 percent with it [3]. SkillGLoW shares the commitment to execution-grounded admission but differs in what it evolves: Recuris patches memory components behind a fixed validation gate, while SkillGLoW re-derives the global layer each round from that round's local evidence and keeps no utility scorer [1][3].

SCAFFOLD offers the strongest competing account of where gains come from. Its ablation attributes the largest single drop, 8.0 points on WebArena, to removing distillation into the base policy, and 6.3 points to removing recursive composition, with MDL compaction contributing a modest 3.1 points while keeping the library bounded [8]. That ordering is a direct challenge to a purely prompt-level library: SCAFFOLD's authors argue a prompt-only skill library, however well structured, is fundamentally limited by the frozen base policy [8]. SkillGLoW does not test distillation, so its 17.2-point gain is attributable to organization and admission alone, and the two explanations are not separated by the supplied evidence [1][8]. SkillZip sits between them as a compression-layer alternative, preserving operation-to-verifier paths and dependency relations through reversible macros with hydration on demand [2].

Transfer to unseen tasks, and the difference between procedure and task memory

The transfer evidence is the paper's cleanest argument that what consolidates is procedure rather than task memory. Injecting the library the training stream produced, unmodified, into 60 unseen ALFWorld tasks lifts mean success from 73.9 to 83.9 percent across three models, with the smallest gain going to the model with the highest baseline, consistent with a ceiling effect [1]. On software repair the same library lifts the resolve rate on 30 unseen instances from 40.0 to 45.6 percent for MiniMax-M3 as a mean of three trials, the workload where transfer should be hardest because every instance is a different repository [1]. The paper contrasts this with a reported 3.7 percent average gain for cross-domain memory transfer in a setting that pools heterogeneous domains rather than holding out instances of one benchmark [1].

The family structure itself is finer-grained than the dataset's taxonomy by design. On ALFWorld, clustering produced K=15 families against 6 canonical task types, with purity 0.950 and ARI 0.605, and the gap comes from over-segmentation rather than mixing types; removing the 5 singleton families leaves the remaining 35 tasks at purity 0.943 [1]. Two families cut across the goal taxonomy along a shared procedure, for example an egg cooled then placed in the microwave and an egg heated then placed in the fridge, where appliance and receptacle are swapped but the procedural skeleton is identical [1]. The terminal mapping shows the same selectivity in the other direction: nine repairs against one pass-to-fail, and a byte-scanning prior that repaired only the two tasks actually solved by scanning raw bytes while leaving three others untouched [1].

Where the conclusion stops: verifier dependence, scale, and untested inheritance

The stated boundary is explicit: conclusions are limited to four benchmarks and three models, the commit gate depends on real execution, and universal open-domain long-term deployment is not demonstrated [1]. Transfer was shown where task categories recur, and whether a prior survives a genuine domain change remains open, as does cross-model library inheritance, since a prior is plain text but no such handoff was tested [1]. The library-size claim is also bounded by the family structure rather than the stream, so the 3.6x compression ratio reflects how many distinct procedures were found on these four streams, not a general law [1].

Independent evidence sharpens rather than resolves the boundary. SkillFlow, evaluating lifelong skill discovery across 11 model variants and four agent harnesses, finds selective rather than universal gains, with stronger stacks converting experience into compact reusable procedures while weaker ones show a creation-reuse coordination gap and fragmented libraries, and it characterizes failure as fragmentation through skill inflation rather than skill scarcity [7]. Skill retrieval work at the scale of roughly one million distinct skills finds that retrieval alone does not solve scalable augmentation, because agents still fail to recognize when external help is needed and which retrieved skills are worth loading [5]. A survey of dynamic agent skills similarly notes that systems often fail by skill inflation rather than inability to write skills, and that a principled policy for what enters the library is missing [6]. SkillOptimizer approaches the same target from the opposite direction, optimizing a decomposed skill document against criteria fit on synthetic prompts rather than scalar task reward [4], which is a different admission signal than SkillGLoW's verifier-grounded gate and has not been compared head to head in the supplied material.

About These Sources

This research page is built on 8 studies (7 peer-reviewed, 1 preprint) — published in 2026, 6 from 2024 or later, collectively cited 64 times — selected as the most relevant from 8 studies that passed quality screening, drawn from 52 papers retrieved from a database of over 500 million.

Sources used in this answer