LLM agents retrieve skills from public hubs that hold up to about 100,000 entries and serve many models. As models improve, a hub must find the skills a given model has internalized, meaning it can do without them, and retire them for that model, yet internalization has not been measured with known reliability. Published rules that compare runs with and without a skill retire on the gain alone. The gain is near zero both for an internalized skill and when neither the model nor the skill delivers the capability, a state we call uncovered. We instead withhold the skill, score the model on a task synthesized from it, and trust a high unaided score only when another model gains from the skill, since nearly half of high scores come from tasks no model benefits from. Under four models, a published gain rule applied with one task per skill removes over half of the uncovered skills. On SkillsBench, 28 of its 31 retirements rest on probes no run passed in full, while retiring on the unaided score costs no detectable reward. Unaided verdicts reproduce across tasks more often than the gain rule's (Cohen's kappa 0.42 against 0.23) but differ across models, so a hub must read them per model. We release 1,251 skills with tasks and item scores under four models.
Figures & tables
Figure 1. The measurement pipeline for one skill. Panel 3 lists the item scores behind U and A , with a dropped generic item in grey. Panel 4 gives this skill’s state, the share of uncovered skills that the band rule ∣A−U∣<0.10 and the level rule retire over the four models, and a second skill whose level-rule verdict differs across models (DS-V4 is DeepSeek-V4-Flash).
Figure 2. The gated adversarial library by unaided and aided score. The band retires the coloured skills between the dotted lines, the uncovered ones in red, and the level rule retires those right of the vertical line. Marker area grows with the number of skills at a point.
Adversarial
Tessl-single
Rule
retired
uncovered share
uncovered removed
internalized removed
retired
uncovered share
uncovered removed
internalized removed
gain <0
9.4%
11.4%
12.3%
10.7%
2.3%
50.0%
19.0%
2.7%
gain ≤0
25.0%
18.5%
53.4%
45.4%
4.8%
35.3%
28.6%
12.3%
∣ gain ∣<0.10
27.8%
17.9%
57.5%
61.3%
15.0%
18.9%
47.6%
45.2%
U≥0.8
26.4%
0
0
77.1%
13.8%
0
0
67.1%
Table 1. What each retirement rule removes, with Qwen3.6-27B as target.
Method
Setting
Retires when
Guard
SKILL0 ( Lu et al., 2026 )
RL training
gain ≤0 , budget
none
SKILLC ( Lin et al., 2026 )
RL training
smoothed gain <0 , budget
none
SLIM ( Shen et al., 2026 )
RL training
smoothed LOO gain
exposure, patience
ASSAY ( Wang et al., 2026c )
library
∣ masked gain ∣<0.10
split by spread
Library Drift ( Zhang et al., 2026b )
library
aided net success ≤−0.10
100 trials
SkillOps ( Pu et al., 2026 )
library
aided utility < threshold
duplicate exists
Table 2. Retirement criteria in the literature and in this paper. LOO is leave-one-out, RL is reinforcement learning, and budget is a cap on retained skills that falls to zero.
Uncovered removed
Internalized removed
Target
Base
band
binary
band
level
Qwen3.6-27B
8.6%
57.5%
79.5%
61.3%
77.1%
Qwen3.5-397B
9.4%
62.8%
98.7%
57.9%
72.7%
DeepSeek-V4-Flash
15.9%
65.4%
91.7%
62.1%
75.3%
Qwen3.8-27B
12.8%
58.9%
86.9%
65.9%
80.8%
Table 3. The audit with each model in turn as target, on 844, 834, 839 and 836 gated skills. Base is the uncovered base rate.
Figure 3. Reproducibility across two tasks of a skill, on the 124 skills scored on all task-completion items of both. Left, parallel-forms reliability. Right, κ of each rule’s verdict, with raw agreement above each bar and 95% bootstrap intervals.
Rule
retired
tasks
slots
shared
Δ
interval
gL
18
17
18
—
−0.2
[−2.7,2.4]
gBbin
31
25
29
10
+0.6
[−1.7,3.1]
gBall
15
14
14
7
−0.5
[−2.0,0.7]
gB∗
18
17
17
9
−0.8
[−3.2,1.7]
Table 4. Retirements on the 59 admitted probes that give no item away, with the benchmark tasks and skill slots they touch, the overlap with gL and the cost Δ of Eq. ( 14 ) in points of mean reward. Target Qwen3.6-27B, at least five runs per touched task for each deletion, nine for the full library.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Attempted
3,000
Scored unaided once
1,323
too easy
−859
Both arms, three runs
1,251
not performable
−708
Score computable
1,165
generation error
−34
Passes validity gate
844
Appendix
Table 5. Dataset construction counts.
Base rate
Bound.
ungated
gated
Tessl
in band
enrich.
retired
0.3
8.7%
5.6%
2.8%
13.6%
2.45 ×
68.1%
0.4
11.8%
8.6%
5.9%
17.9%
2.07 ×
57.5%
0.5
14.1%
11.1%
7.1%
19.6%
1.76 ×
48.9%
0.6
17.9%
15.4%
9.3%
24.3%
1.57 ×
43.8%
0.7
23.3%
21.4%
15.8%
27.7%
1.29 ×
35.9%
Appendix
Table 6. The band audit across the high–low boundary. The last three columns use the gated adversarial set.
Adversarial
Tessl-single
τ
retired
precision
recall
retired
precision
recall
0.7
34.7%
92.5%
100%
20.6%
100%
100%
0.8
26.4%
93.7%
77.1%
13.8%
100%
67.1%
0.9
18.6%
93.6%
54.2%
5.9%
100%
28.8%
1.0
11.8%
94.0%
34.7%
1.4%
100%
6.8%
band
27.8%
70.6%
61.3%
15.0%
62.3%
45.2%
Appendix
Table 7. The level rule across its threshold τ , with the band.
State
Skills
gain <0
gain ≤0
band
band, A≥0.7
level
Internalized
271
29
123
166
166
209
Skill works
198
0
0
0
0
0
Uncovered
73
9
39
42
0
0
Skill hurts
7
7
7
0
0
6
Between thresholds
295
34
42
27
1
8
Appendix
Table 8. Skills of each state retired by each rule.
Skills have emerged as a practical and effective approach for enhancing LLM agents at inference time through structured packages of knowledge. However, existing evaluations largely measure whether skills improve aggregated task success, leaving a more fundamental question underexplored: When do skills help, why do they work, and where do they fail? Through controlled experiments across various benchmarks, agent harnesses and LLMs, we isolate the effects of representation, outcome annotation, retrieval difficulty, and cross-framework robustness of skills. To further answer this question, we design a contrastive study that combines controlled quantitative experiments with paired trajectory analysis. We normalize 8,135 trial records from controlled experiments and retain 238 valid unique labels from 240 open-coded records. We consolidate these observations into a taxonomy of three high-level categories and twelve skill-use modes: skills work when noisy trajectories become procedural anchors that stabilize execution. Skills improve over Workflow Memory by 6.06 points in matched comparisons. Procedural anchoring accounts for 65.7% of skill cases, versus 4.5% for explicit knowledge injection, showing that skills stabilize action rather than inject missing facts. Retrieval is a separate bottleneck: as pools grow from 5 to 100, actual-use precision falls from 29.6% to 3.3%. Confusable distractors impair offline identification, yet downstream success remains stable; exact ground-truth invocation is neither sufficient nor necessary. Skills fail under brittle assumptions, incompatible contexts, or insufficient adaptation. These findings move evaluation beyond aggregate success rates and guide reliable self-evolving agents.
Zhiyuan Jiang, Fangrui Huang, Hanwen Xing +6
Princeton University · Stanford University · University of Southern California +2
Skill documents, structured natural-language instructions that guide Large Language Model (LLM) agents, are critical to modern agent frameworks, yet LLMs struggle to write skills that actually work. On SkillsBench, human-authored skills improve pass rates by 16.2 percentage points, while LLM-authored skills provide no measurable gain. We introduce SkillAxe, a fully unsupervised framework that enables LLMs to iteratively diagnose and refine their own skills. SkillAxe decomposes skill quality into four interpretable dimensions (quality impact, trigger precision, instruction compliance with fault attribution, and solution-path coverage), producing structured improvement briefs that require no ground-truth labels, test suites, or environment rewards. On SkillsBench, SkillAxe improves pass rates by 28% relative over unimproved LLM skills and closes 47--67% of the gap to human-authored skills. We validate the approach as a continuous improvement engine in the wild on SpreadsheetBench, where a SkillAxe-built skill library learns from past agent trajectories and raises pass rate from 16.0% to 52.0% using only 22 skills.
Large language model (LLM) agents can induce skills from completed tasks and reuse them later to grow more capable with experience. In practice, induced skills may transfer unreliably and can even harm the agent that retrieves them. When agent-induced skills transfer reliably across tasks remains an open question. We conduct a comprehensive and controlled study of how the way skills are induced shapes their transfer across tasks. Specifically, we compare task-level with subtask-level skill induction and text with code skill formats, the two axes along which existing methods differ. Task-level skills mostly reduce the agent's performance below its no-memory baseline while subtask-level skills raise it above on average, and text skills transfer better than code skills. To further understand our findings, we examine two complementary properties of the induced skills: specificity, which measures how closely a skill matches real tasks, and abstractness, which measures how evenly its relevance spreads across tasks. Neither property alone predicts task success, but their combined effect does, which we propose as a skill utility score. The score correlates consistently with task success when skills are transferred, and subtask-level and text skills score higher. Computing skill utility only needs the skills and task descriptions, so our score serves as a lightweight diagnostic of a skill memory for practitioners.