Video yükleniyor...
Video Yüklenemedi
We present Tropical Reinforcement Learning. Standard RL for LLMs cannot learn compositions effectively, for a fundamental reason: its algebra is not optimal. It sums the probabilities of every successful answer, measuring how often a model succeeds, not which solution succeeds.
43,127 görüntüleme • 1 ay önce •via X (Twitter)
22 Yorum

Our fix: change the algebra. Instead of summing over solutions, take the max → the tropical (max-plus) semiring. Each state's value is now the log-prob of its best verified solution, with the path itself. Good prefixes and suffixes from different rollouts can be joined.

Results: TROPIC beats the strongest on-policy baselines by up to 16 percentage points across four benchmarks. Big picture: changing the algebra of RL, not just its estimators, can make LLMs much better at compositional reasoning.

Paper: Code:

This is insane clickbait 🙂 I thought SSI released something

Sorry, I didn’t even think someone would see it that way.

It's no problem! Nice work, tho!

Very similar to our agentic discovery on sudoku/arc. Although our focus was max likelihood not RL, the idea is the same: backprop through the winning head but transfer state to losers too so rich don't get richer.

This is exactly right. Well trained means to represent the underlying algebra precisely.

if this holds, a lot of the pass@k shrinkage people pin on rl is really just the sum being the wrong thing to optimize

More results on pass@k soon.

looking forward to it. does the gap keep widening as k grows or does it level off past some point?

Fun fact: the tropical semiring was named such because it came from Brazil and the French mathematicians who saw just thought it was wild they were getting mathematics out of Brazil.

Yeah, max-plus is used more often, but I prefer tropical

It makes the LLM go bananas

😂😂

Finally, congrats on the release! 🎉

@tairov you had me at Ilya and had me through the end

and how would doing the RL in panama change any of this

Electricity is cheap in Panama :)

Ilya did not invent GPT-4 or GPT-3, it was work done by other people in OpenAI

Replacing sum with max plus turns the objective into a clear best path EV, which explains the 16% lift you report

Summing probability over successes means two half-right traces outweigh one full-right one. Max fits pass@1 but kills gradient on near-misses, which teach the most. Fix is a temperature'd softmax, not a hard max.
