SCATR paper accepted to COLM 2026 workshop, training lightweight decision models in the spirit of Jev
The authors present SCATR, which trains a lightweight scorer on a small calibration set to rank candidate answers, a direction they say aligns with Jev's specialized fast decision models, improving over confidence-based ranking by up to 9% on coding and math tasks.
Happy to share that SCATR has been accepted to the Workshop on Efficient Reasoning at #COLM2026! 🎉 One thing we’ve been thinking about a lot is what happens as test-time compute and agentic systems generate more and more candidate solutions. At that point, generation is only half the problem. The other half is: Which answer should you trust, and how do you make that decision cheaply? 🤔⚡ That’s also why we find the direction behind systems like Jev interesting: not every decision needs another expensive autoregressive model. Sometimes, you want a specialized, fast, calibrated decision model. SCATR can be viewed as a simple recipe for training exactly that kind of model. Given an LLM and a small calibration set, it learns a lightweight scorer over the model’s own representations to rank candidate answers, without a heavyweight verifier. 🎯 Across coding and math, SCATR improves over confidence-based ranking by up to 9%, while using dramatically fewer trainable parameters than heavier alternatives. 🚀 Generate with powerful models. Train specialized models to make the decisions. 🧠➡️🎯 We’ll be presenting SCATR in person at COLM at the Workshop on Efficient Reasoning, come say hi! 👋 🎥 Quick overview below 📄 https://arxiv.org/abs/2604.16535
This post quotes another post: x.com/thecekbote/status/2046715476018074095