The authors trained a shared Qwen3 policy to choose between thinking and direct-answer modes for each prompt.
Models learn when to reason through counterfactual routing credit
CounterRoute trains one policy to select between deliberate reasoning and direct answers while separately improving responses within each mode.
Big Tech
Amazon.com
Research Digest··2 min read
Jiao et al.
Why this paper
From Amazon.com
In one line
CounterRoute jointly learns when language models should reason or answer directly, improving benchmark accuracy while reducing generated tokens versus always-thinking Qwen3 checkpoints.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ✓Compute or model size stated (params 8B)
- ·No stated limitations found
- ✓Reports numbers on named benchmarks
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§