Loading video...

Video Failed to Load

Go Home

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 views • 1 month ago •via X (Twitter)

22 Comments

Arip's profile picture
Arip1 month ago

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.

Arip's profile picture
Arip1 month ago

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.

Arip's profile picture
Arip1 month ago

Paper: Code:

Swaroop Nath's profile picture
Swaroop Nath1 month ago

This is insane clickbait 🙂 I thought SSI released something

Arip's profile picture
Arip1 month ago

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

Swaroop Nath's profile picture
Swaroop Nath1 month ago

It's no problem! Nice work, tho!

Joshua V. Dillon's profile picture
Joshua V. Dillon1 month ago

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.

Ryan McCormick's profile picture
Ryan McCormick1 month ago

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

Saksham Jain's profile picture
Saksham Jain1 month ago

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

Arip's profile picture
Arip1 month ago

More results on pass@k soon.

Saksham Jain's profile picture
Saksham Jain1 month ago

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

Catch Vartan's profile picture
Catch Vartan1 month ago

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.

Arip's profile picture
Arip1 month ago

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

Henrique Dalposso's profile picture
Henrique Dalposso1 month ago

It makes the LLM go bananas

Arip's profile picture
Arip1 month ago

😂😂

Greer's profile picture
Greer1 month ago

Finally, congrats on the release! 🎉

Carter Leffen's profile picture
Carter Leffen1 month ago

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

bitflipgremlin's profile picture
bitflipgremlin1 month ago

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

Arip's profile picture
Arip1 month ago

Electricity is cheap in Panama :)

DRAMYYDS (全网吹芯第一人)'s profile picture
DRAMYYDS (全网吹芯第一人)1 month ago

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

Robin | Poker x AI's profile picture
Robin | Poker x AI1 month ago

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

Coffee Cat's profile picture
Coffee Cat1 month ago

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.

Related Videos