Raina T. Wu

Hi! I’m Raina, an undergraduate at MIT studying computer science (expected to graduate May 2027). I work on reinforcement learning and post-training for LLM reasoning, and more broadly on RL for generative models.

Most recently, I was a research intern at IBM Research, where I ran ablations and reproductions of self-distillation pipelines for LLM reasoning, characterized their failure modes, and studied where privileged-teacher signal localizes within a rollout across Qwen3 models. Before that, I worked on RL fine-tuning of flow models at the MIT-IBM Watson AI Lab, sketch-to-3D retrieval for generative design at the MIT DeCoDE Lab, and attestation for AI agents negotiating inside trusted execution environments at the MIT Media Lab. I got my start in research through MIT PRIMES-USA, studying influence prediction in collaboration networks.

Over the past few summers, I’ve interned at Talkdesk (multi-agent systems) and Hudson River Trading (WITTI).

I’m currently a TA for 18.404 Theory of Computation. In high school I did competition math (MOP, USAJMO winner, Math Prize for Girls top 12), and these days I still take the Putnam for fun (top 300). I also sing in MIT Syncopasian, one of MIT’s many a cappella groups.

See my projects for details on my research and coursework, or my CV for the full picture.

news

Sep 02, 2026	TAing 18.404 Theory of Computation this fall.
Sep 01, 2026	Submitted our prefix-advantage study of self-distillation to the NeurIPS 2026 workshop on Transitioning from Pre-Training to Post-Training.
Aug 01, 2026	Tailored primitive initialization is the secret key to reinforcement learning was published at ACL 2026.
Jun 01, 2026	Started as a research intern at IBM Research, working on RL and self-distillation for LLM reasoning.
Jan 05, 2026	Spent January at Hudson River Trading as a WITTI wintern.

selected publications

ACL
Tailored primitive initialization is the secret key to reinforcement learning

Y. Yao, G. Zeng, R. Wu, and 4 more authors

In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 2026

Abs DOI Bib HTML

Introduces Tailor, a warm-start strategy that initializes RL fine-tuning of LLM reasoners from a tailored set of reasoning primitives. I built the embedding-similarity pipeline used for the paper’s diversity analysis, showing that Tailor’s warm-start data is more reasoning-diverse than Standard-CoT and 4-STaR baselines.
@inproceedings{yao2026tailored, title = {Tailored primitive initialization is the secret key to reinforcement learning}, author = {Yao, Y. and Zeng, G. and Wu, R. and Zhang, Y. and Zhao, D. and Hong, Z.-W. and Gan, C.}, booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)}, pages = {33300--33318}, address = {San Diego, CA}, year = {2026}, doi = {10.18653/v1/2026.acl-long.1537}, }
Physica A
Influence prediction in collaboration networks: An empirical study on arXiv

M. Lin, L. P. Schaposnik, and R. Wu

Physica A: Statistical Mechanics and its Applications, 2026

Abs DOI Bib HTML

Combines path-based link prediction with influence-maximization heuristics to identify future vital nodes in weighted collaboration networks, evaluated on arXiv co-authorship data. Introduces the RA-2 and quasi-local RA-2 similarity metrics. Authors are listed alphabetically.
@article{lin2026influence, title = {Influence prediction in collaboration networks: An empirical study on {arXiv}}, author = {Lin, M. and Schaposnik, L. P. and Wu, R.}, journal = {Physica A: Statistical Mechanics and its Applications}, volume = {689}, pages = {131451}, year = {2026}, doi = {10.1016/j.physa.2026.131451}, }