Video yükleniyor...

Video Yüklenemedi

Ana Sayfaya Dön

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

Arip profil fotoğrafı
Arip1 ay önce

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 profil fotoğrafı
Arip1 ay önce

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 profil fotoğrafı
Arip1 ay önce

Paper: Code:

Swaroop Nath profil fotoğrafı
Swaroop Nath1 ay önce

This is insane clickbait 🙂 I thought SSI released something

Arip profil fotoğrafı
Arip1 ay önce

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

Swaroop Nath profil fotoğrafı
Swaroop Nath1 ay önce

It's no problem! Nice work, tho!

Joshua V. Dillon profil fotoğrafı
Joshua V. Dillon1 ay önce

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 profil fotoğrafı
Ryan McCormick1 ay önce

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

Saksham Jain profil fotoğrafı
Saksham Jain1 ay önce

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 profil fotoğrafı
Arip1 ay önce

More results on pass@k soon.

Saksham Jain profil fotoğrafı
Saksham Jain1 ay önce

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

Catch Vartan profil fotoğrafı
Catch Vartan1 ay önce

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 profil fotoğrafı
Arip1 ay önce

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

Henrique Dalposso profil fotoğrafı
Henrique Dalposso1 ay önce

It makes the LLM go bananas

Arip profil fotoğrafı
Arip1 ay önce

😂😂

Greer profil fotoğrafı
Greer1 ay önce

Finally, congrats on the release! 🎉

Carter Leffen profil fotoğrafı
Carter Leffen1 ay önce

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

bitflipgremlin profil fotoğrafı
bitflipgremlin1 ay önce

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

Arip profil fotoğrafı
Arip1 ay önce

Electricity is cheap in Panama :)

DRAMYYDS (全网吹芯第一人) profil fotoğrafı
DRAMYYDS (全网吹芯第一人)1 ay önce

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

Robin | Poker x AI profil fotoğrafı
Robin | Poker x AI1 ay önce

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

Coffee Cat profil fotoğrafı
Coffee Cat1 ay önce

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.

Benzer Videolar