Загрузка видео...

Не удалось загрузить видео

На главную

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 просмотров • 1 месяц назад •via X (Twitter)

Комментарии: 22

Фото профиля Arip
Arip1 месяц назад

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.

Фото профиля Arip
Arip1 месяц назад

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.

Фото профиля Arip
Arip1 месяц назад

Paper: Code:

Фото профиля Swaroop Nath
Swaroop Nath1 месяц назад

This is insane clickbait 🙂 I thought SSI released something

Фото профиля Arip
Arip1 месяц назад

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

Фото профиля Swaroop Nath
Swaroop Nath1 месяц назад

It's no problem! Nice work, tho!

Фото профиля Joshua V. Dillon
Joshua V. Dillon1 месяц назад

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.

Фото профиля Ryan McCormick
Ryan McCormick1 месяц назад

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

Фото профиля Saksham Jain
Saksham Jain1 месяц назад

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

Фото профиля Arip
Arip1 месяц назад

More results on pass@k soon.

Фото профиля Saksham Jain
Saksham Jain1 месяц назад

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

Фото профиля Catch Vartan
Catch Vartan1 месяц назад

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.

Фото профиля Arip
Arip1 месяц назад

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

Фото профиля Henrique Dalposso
Henrique Dalposso1 месяц назад

It makes the LLM go bananas

Фото профиля Arip
Arip1 месяц назад

😂😂

Фото профиля Greer
Greer1 месяц назад

Finally, congrats on the release! 🎉

Фото профиля Carter Leffen
Carter Leffen1 месяц назад

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

Фото профиля bitflipgremlin
bitflipgremlin1 месяц назад

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

Фото профиля Arip
Arip1 месяц назад

Electricity is cheap in Panama :)

Фото профиля DRAMYYDS (全网吹芯第一人)
DRAMYYDS (全网吹芯第一人)1 месяц назад

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

Фото профиля Robin | Poker x AI
Robin | Poker x AI1 месяц назад

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

Фото профиля Coffee Cat
Coffee Cat1 месяц назад

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.

Похожие видео