The authors examined IFBench, which adds programmatically checkable output constraints to user requests.
Calibrated rewards curb answer degradation during verifiable instruction training
CaLIF supplements constraint-verification rewards with bounded response-likelihood signals, helping models follow formatting constraints without sacrificing substantive answers.
Top University
Sangyeon Yoon · Wooseok Seo · Hamish Ivison · Hyesoo Hong · Donghoon Kim · Wonjun Kang · +2 more
Yonsei University · University of Washington · FuriosaAI
Research Digest··3 min read
Yoon and colleagues show that programmatic instruction-following verifiers can award full credit to responses that satisfy output constraints while barely addressing the underlying request.
Why this paper
From University of Washington and 2 others
In one line
Calibrating likelihood against a reference policy reduces reward hacking in verifiable instruction following.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·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.
§