Agent frameworks increasingly package procedural knowledge as skills: instruction files an agent reads on demand, while public libraries now hold thousands of them. Which skill to read has thus become a decision the policy itself makes in the middle of an episode, yet no existing signal trains it. We show that the default remedy, outcome-rewarded RL over the candidate slate, cannot teach it, for a structural reason we identify and name selector credit starvation: under a broadcast, sequence-level advantage, the few tokens that name the chosen skill carry a vanishing share of the loss, and the credit they inherit is increasingly wrong-signed as trajectories lengthen. A correct choice is punished whenever the execution after it fails, even though the choice itself is among the most valuable decisions in the trajectory. Auditing a completed run's own training artifacts confirms all three properties, each worsening monotonically with horizon. SkillGate removes the failure by construction: it partitions the token support into two disjoint credit channels, outcome credit reaching only execution tokens, and a separate action-local advantage reaching exactly the skill-naming tokens, positive only when a trajectory's single read is the correct one. On five agentic benchmarks under a 16-candidate slate, SkillGate lifts a 9B policy from 40.8% to 53.2% trial success, well ahead of the identical budget spent on outcome reward alone, while cutting exposure to misleading candidates by two thirds and reading fewer skills.
Figures & tables
Figure 1: Selector credit starvation. One broadcast advantage updates both the few tokens choosing a skill and the thousands executing the task ( top ); on 12,800 training trajectories, the choice’s loss share dilutes 7× with length ( Share ), its credit is increasingly wrong-signed ( Sign ), yet matched prompt groups show a +11.2 pp success gap for the correct read ( Value ).
Figure 2: SkillGate overview. The task channel group-normalises terminal outcomes into Atask and broadcasts it over execution tokens with the entire read call cut out; the selector channel scores read actions with the clean single-oracle utility and places Asel on the identity tokens alone. The supports are disjoint and each channel’s token weights sum to N before λ is applied; both enter one GRPO update.
Trial success (%)
Read behaviour (%)
Method
Claw-Eval
SkillsBench
SETA
SWE
TB2
Overall
Oracle ↑
Mislead. ↓
Frontier models (capability ceiling, not controlled)
DeepSeek-V4-Flash
70.2
15.6
66.7
55.0
46.9
61.0
40.8
23.9
GLM-5
75.8
15.6
59.2
45.0
56.2
60.8
39.0
26.2
DeepSeek-V4-Pro
70.8
12.5
60.8
55.0
62.5
60.5
38.2
20.8
Kimi-K2.6
71.4
12.5
56.7
55.0
71.9
60.3
45.7
21.0
Table 1: Trial success (%) on the standard mixed slate, and read behaviour. Oracle / misleading : fraction of trials reading at least one such skill (a trial can read both). Best 9B-scale per column in bold .
Design
Where the credit lands
Trial succ. (%)
Clean single- oracle (%)
Oracle (%)
Mislead. (%)
Reads/ trial
SkillRL (outcome only)
nowhere (anchor)
42.1
21.4
54.3
69.6
1.88
Group-level regret
the prompt group
41.8
15.7
33.6
47.5
1.23
Trajectory bonus
the whole trajectory
41.8
33.9
47.9
55.4
1.33
Action credit
first oracle read
45.0
64.6
80.0
32.5
1.26
SkillGate
the only read, if oracle
50.0
75.4
83.9
21.8
1.11
Table 2: Credit-design ablation, 280-trial protocol: five placements of the selector signal from the same initialisation and steps. Clean single-oracle : exactly one attributed read and it is the oracle. Best results in bold .
Figure 3: The five credit designs of \cref tab:ablation: task outcome (a), what gets read (b), and how much gets read (c). Only credit that reaches the identity tokens changes the choice, and the single-read rule converts it into success.
Setting
Trial succ. (%)
Top-1 oracle † (%)
Oracle (%)
Mislead. (%)
Reads/ trial
SFT, standard mixed slate
37.1
—
37.9
61.8
1.34
SFT + “read only one” prompt
35.0
—
41.4
59.3
1.35
SFT-9B router → SFT
40.7
60.0
57.5
36.8
0.94
Qwen3.5-27B router → SFT
36.8
68.6
68.2
24.3
0.95
Qwen3 reranker top-1 → SFT
31.8
27.1
24.6
60.4
0.89
SkillGate , standard mixed slate
50.0
78.9
83.9
21.8
1.11
Table 3: External selection, 280-trial protocol. Routers advertise one chosen skill to the frozen SFT executor; oracle-only injection is a ceiling that requires knowing the answer. † Top-1 is over 70 routing decisions for routers but SkillGate’s first attributed read over 280 trials — not directly comparable.
Figure 4: Inference cost, 280-trial protocol. (a) SkillGate vs. SkillRL (outcome only): reads, turns and token usage. (b) Skill-body tokens loaded: measured on-demand reading vs. preloading k candidates’ SKILL.md bodies.
Figure 5: Offline audit of SkillRL (outcome only)’s 12,800 training trajectories, by trajectory-length bin. (a) Median loss-weight share of the read call and of the skill-naming tokens. (b) Within-group signal-to-noise of the read-vs-not advantage gap. (c) Share of oracle reads inheriting a negative advantage; SkillGate’s zero is by construction.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Claw (161)
SB (32)
SETA (120)
SWE (40)
TB2 (32)
All (385)
Qwen3.5-9B
72
0
29
0 6
0 3
110
Qwen3.5-27B
87
3
51
16
11
168
SFT (RL init)
82
2
48
18
0 7
157
Selection BC
84
5
52
20
11
172
SelSkill-DPO
85
0
57
24
12
178
Skill-free RL
90
3
57
17
10
177
Appendix
Table 4: Raw successes for the non-frontier rows of \cref tab:main. Column headers give the fixed number of trials for each benchmark.
Method
pass@1 (%)
Grader mean
Read oracle (%)
Read misleading (%)
Qwen3.5-9B
42.9
0.562
0 7.5
0 6.8
Qwen3.5-27B
53.1
0.672
13.6
0 4.1
SFT (RL init)
51.7
0.647
42.9
42.2
Selection BC
51.7
0.641
71.4
23.8
SelSkill-DPO
53.7
0.675
59.2
35.4
Skill-free RL
56.5
0.696
38.1
51.0
Appendix
Table 5: Results on the 147 held-out Claw tasks, with one trial per task and the standard mixed slate. Grader mean is the benchmark’s partial-credit score. Oracle and misleading reads are overlapping events and can sum past 100%.
Method
Any read (%)
Oracle (%)
Misleading (%)
P(oracle∣read)
Reads/trial
Qwen3.5-9B
15.0
0 5.7
0 8.9
38.1
0.19
Qwen3.5-27B
15.7
0 5.7
0 8.9
36.4
0.18
SFT (RL init)
94.6
37.9
61.8
40.0
1.34
Selection BC
98.2
71.4
31.4
72.7
1.05
SelSkill-DPO
96.4
66.1
51.8
68.5
1.59
Skill-free RL
80.0
35.4
55.0
44.2
1.25
Appendix
Table 6: Read behaviour on the 280-trial protocol. Oracle and misleading exposure are multi-label events and can sum past 100%; reads/trial counts distinct skill names. The untrained rows rarely read, so their low misleading exposure reflects abstention rather than selectivity.
Figure 6: SkillGate minus SkillRL (outcome only) per benchmark on the repeated-trial protocol. Markers are observed differences and bars are 95% task-clustered bootstrap intervals. “All” is the pooled result, not a sixth benchmark.