Reasoning-Finetuning Repurposes Latent Representations in Base Models

Chain-of-thought prompting elicits reasoning in large language models

Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al · 2022

Earlier work this paper cites.

Refusal in language models is mediated by a single direction

Original

Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N · 2024

Earlier work this paper cites.

Steering llama 2 via contrastive activation addition

Rimsky, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A · 2024

Earlier work this paper cites.

Demystifying long chain-of-thought reasoning in llms

Original

Yeo, E.-H., Lin, Z., Han, T., Lin, M., Wang, J., Wang, J., Zhang, J., and Zhuo, D · 2024

Earlier work this paper cites.

I have covered all the bases here: Interpreting reasoning features in large language models via sparse autoencoders

Original

Galichin, A., Dontsov, A., Druzhinina, P., Razzhigaev, A., Rogov, O. Y., Tutubalina, E., and Oseledets, I · 2025

Cited alongside, same era.

Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

Original

Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al · 2025

Cited alongside, same era.

s1: Simple test-time scaling

Original

Niklas Muennighoff, Zitong Yang, W. S. X. L. L. L. F.-F. H. H. L. Z. P. L. E. C. T. H · 2025

Cited alongside, same era.

Understanding reasoning in thinking language models via steering vectors

Venhoff, C., Arcuschin, I., Torr, P., Conmy, A., and Nanda, N · 2025

Closest in time.

Step back to leap forward: Self-backtracking for boosting reasoning of language models

Original

Yang, X.-W., Zhu, X.-Y., Wei, W.-D., Zhang, D.-C., Shao, J.-J., Zhou, Z., Guo, L.-Z., and Li, Y.-F · 2025

Closest in time.

On the emergence of thinking in llms i: Searching for the right intuition

Original

Ye, G., Pham, K. D., Zhang, X., Gopi, S., Peng, B., Li, B., Kulkarni, J., and Inan, H. A · 2025

Closest in time.