Didactic to Constructive: Turning Expert Solutions into Learnable Reasoning
Ethan Mendes, Jungsoo Park, Alan Ritter · 2026
Read Didactic to Constructive: Turning Expert Solutions into Learnable Reasoning by Ethan Mendes, Jungsoo Park, Alan Ritter (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-02405.txt.
Description
Ethan Mendes and 2 co-authors teach language models from expert human solutions that models cannot yet reach on their own, in a 2026 arXiv paper. Hard problems often yield no correct sampled answer to reinforce, and no stronger model exists to copy. Expert solutions exist, but plain imitation fails because experts write for humans, skipping steps a model needs.
Their method, distribution aligned imitation learning, works in two steps. First, expert solutions are rewritten into detailed reasoning traces that match the model's own style. Then a contrastive objective, which pushes the model toward expert moves and away from its defaults, focuses learning on the expert's insights. With fewer than a thousand expert solutions, reasoning scores rose sharply, efficiency doubled, and gains transferred to new domains.
Catalog record
| CREATOR | Ethan Mendes, Jungsoo Park, Alan Ritter |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-02405