Training Small Language Models for Transferable Tool Selection
3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked
A small open model, Qwen3-8B, is post-trained with reinforcement learning to select tools for multi-step instructions in a Plan-Retrieve-Select loop, where a retriever is invoked on demand. On ToolBench it outperforms the much larger Llama-4-Maverick (402B) with prompt tuning.
Score
0.66total under rubric v2.0
- Significance0.59
- Reader interest0.70
- Audience breadth0.59
- Practical relevance0.70
- Substantiated0.82
- Technical depth0.88
- New information0.91
- Counts against: Promotional0.05
- Counts against: Overclaim0.30
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.67 | 0.40 | 0.72 | 0.59 | 0.32 | +0.25 | +0.149 |
| Reader interest | 0.60 | 0.60 | 0.91 | 0.70 | 0.31 | +0.20 | +0.141 |
| Audience breadth | 0.62 | 0.45 | 0.69 | 0.59 | 0.24 | +0.10 | +0.059 |
| Practical relevance | 0.73 | 0.60 | 0.78 | 0.70 | 0.18 | +0.15 | +0.106 |
| Substantiated | 0.94 | 0.70 | 0.83 | 0.82 | 0.24 | +0.10 | +0.082 |
| Technical depth | 0.95 | 0.75 | 0.94 | 0.88 | 0.20 | +0.10 | +0.088 |
| New information | 0.94 | 0.80 | 0.98 | 0.91 | 0.18 | +0.10 | +0.091 |
| Promotional | 0.05 | 0.10 | 0.01 | 0.05 | 0.09 | −0.15 | −0.008 |
| Overclaim | 0.42 | 0.30 | 0.18 | 0.30 | 0.24 | −0.15 | −0.045 |
| Sum of contributions | +0.662 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.662 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.