Atharv Naphade

Research @ Applied Compute
incredibly famous influencer

[Atharv Naphade @AtharvNaphade]

@yusufksadek @appliedcompute We tried both, found a statistically insignificant difference between the results! On held out tasks, evaluating the chess trained model on math, the base initialized model did slightly (1.5%) better.

[Atharv Naphade @AtharvNaphade]

@willcb Whoops

[Applied Compute @appliedcompute]

Introducing Spork, an efficient post-training recipe which increases our solve-rate on chess puzzles by 100% when compared to our best GRPO setup.

It uses a GRPO warmup, splits task-specific datasets further into partitions, trains specialists, and then uses multi-teacher on-policy distillation. Like a Spork, it leverages gains by simply consolidating specialists.

[Rhythm Garg @rhythmrg]

Start with a brief full-task GRPO warmup, train specialists on random subsets, then distill them via Multi-Teacher On-Policy Distillation (MOPD).

Of course, splitting your task into semantically meaningful subsets yields stronger results, but random subsets actually captured ~2/3 of the gain on chess and still beat full-task GRPO!

Suggests that much of the benefit comes from learning from diverse, independently trained teachers, even without meaningful task decomposition

[Atharv Naphade @AtharvNaphade]

@appliedcompute Really excited for the community to try this on even more tasks and models!

[Sam Denton @samueldenton]

Okay this one is actually a banger. 🍴your way to SotA results...

Pushes way past stand-alone GRPO

[Applied Compute @appliedcompute]

Introducing Spork, an efficient post-training recipe which increases our solve-rate on chess puzzles by 100% when compared to our best GRPO setup.

It uses a GRPO warmup, splits task-specific datasets further into partitions, trains specialists, and then uses multi-teacher on-policy distillation. Like a Spork, it leverages gains by simply consolidating specialists.

[Applied Compute @appliedcompute]

We're sharing a key step towards enabling Continual Learning at Applied Compute.

In On-Policy Self-Distillation, the teacher model supervises training with privileged information: a hint. The training signal depends heavily on hint quality, but manually refining hints across runs is impractical, and isolating their individual impact is difficult. We instead optimize hints automatically using efficient proxies for their eventual training value.