The authors defined relative long-horizon candidate value as a within-prefix measure: the difference in empirical success probability between a candidate token and a reference (the average of plausible candidates).
Reasoning models encode the future value of alternative next tokens
Linear probes decode relative long-horizon candidate value from hidden states and MLP outputs, outperforming probability baselines.
Chinese Tech
Yixin Yang · Yue Wang · Fandong Meng · Zhifang Sui
Peking University · Tencent Inc.
Research Digest··3 min read
Yang et al.
Why this paper
From Tencent Inc. and Peking University · Part of Reasoning Early Stopping, now 3 papers
In one line
Relative long-horizon candidate value is linearly decodable from reasoning model representations and can guide token selection.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§