Pinned Tweet
New paper!
Want to precisely optimize synthetic training data to do practical or even wacky things?
Dataset Policy Gradients get you there, letting you target any differentiable training or post-training metric. We embedded a QR code in GPT-2’s weights using only training data!
Want to precisely optimize synthetic training data to do practical or even wacky things?
Dataset Policy Gradients get you there, letting you target any differentiable training or post-training metric. We embedded a QR code in GPT-2’s weights using only training data!
7
43
242
64,050
Tristan Thrush retweeted
🥐 Hi everyone! FlashAttention's backward pass has the same zero-sum bug
@jon_barron
just fixed in softmax. With FP4 stochastic rounding it's wrong on every single draw. We fixed it in our ICLR'27 submission, CroissAttn and unlock e2e FP4 training.
@jon_barron
just fixed in softmax. With FP4 stochastic rounding it's wrong on every single draw. We fixed it in our ICLR'27 submission, CroissAttn and unlock e2e FP4 training.
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
24
101
1,132
72,306
Follow our course if you want to gain a kind of knowledge that is forbidden elsewhere!
deep-learning-alchemy.github.io
deep-learning-alchemy.github.io
I do not think that Newton would be very proud of himself about trying to make gold out of lead, and worse if he ever tried to teach that at Cambridge... The word "frontier" does not make something worthy being taught at a top university. Professors have a responsibility to truly understand something before teaching it to students.
2
4
110
8,303
Tristan Thrush retweeted
Deep learning is often taught as a sanitized collection of facts and math. But work at the frontier of AI is messy and empirical.
Our new class, Deep Learning Alchemy, aims to teach the art and science of deep learning in the hands-on spirit of Language Modeling From Scratch.
on by running experiments on small LMs, then test it through prediction quizzes: given an unseen code diff, can you predict what happens?
deep-learning-alchemy.github.io
My co-instructor @kothasuhas and I think there's significant demand for classes that treat students as serious learners who want to be assessed for what they’ve learned, and so we're trying a new exam-heavy approach to the class (7 exams, 85% of the grade, with the remaining 15% in-person discussion assessments).
We’ve designed the class so students can use AI agents to run their experiments and do their assignment, but they then have to demonstrate that mastery on exams without access to the assignment, and any partial credit on the exam has to come through citations to the assignment, incentivizing them to know what’s in their write-up.
The same assignments and quizzes are public for anyone to try (lecture recordings will be released through @StanfordOnline with some delay). We have some very fun assignments and quizzes covering both classic topics (like NQM, run-to-run stochasticity, and scaling laws) and new and timely topics like muP or hyperball. Made by some real experts like @TristanThrush and @wen_kaiyue
Try some quizzes below!
Our new class, Deep Learning Alchemy, aims to teach the art and science of deep learning in the hands-on spirit of Language Modeling From Scratch.
on by running experiments on small LMs, then test it through prediction quizzes: given an unseen code diff, can you predict what happens?
deep-learning-alchemy.github.io
My co-instructor @kothasuhas and I think there's significant demand for classes that treat students as serious learners who want to be assessed for what they’ve learned, and so we're trying a new exam-heavy approach to the class (7 exams, 85% of the grade, with the remaining 15% in-person discussion assessments).
We’ve designed the class so students can use AI agents to run their experiments and do their assignment, but they then have to demonstrate that mastery on exams without access to the assignment, and any partial credit on the exam has to come through citations to the assignment, incentivizing them to know what’s in their write-up.
The same assignments and quizzes are public for anyone to try (lecture recordings will be released through @StanfordOnline with some delay). We have some very fun assignments and quizzes covering both classic topics (like NQM, run-to-run stochasticity, and scaling laws) and new and timely topics like muP or hyperball. Made by some real experts like @TristanThrush and @wen_kaiyue
Try some quizzes below!
43
313
2,457
178,037
Presenting at COLM in a few min; poster today from 11 to 1!
New paper!
Want to precisely optimize synthetic training data to do practical or even wacky things?
Dataset Policy Gradients get you there, letting you target any differentiable training or post-training metric. We embedded a QR code in GPT-2’s weights using only training data!
0
1
18
1,218
Tristan Thrush retweeted
Introducing the world's fastest tokenizer implementation, Gigatoken!
Gigatoken is ~500-1000x faster than HuggingFace, and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines.
These baselines are already multithreaded Rust implementations! 🧵
Gigatoken is ~500-1000x faster than HuggingFace, and ~100x faster than OpenAI's tiktoken for most tokenizer definitions on most machines.
These baselines are already multithreaded Rust implementations! 🧵
84
481
4,135
741,609
Tristan Thrush retweeted
Excited to share our new work on training language models to use multiple axes of inference compute — sequential, parallel, and aggregative — end-to-end, led by @jubayer_hamid and @ifdita_hasan.
LLMs already use many forms of compute at test time, so they should learn to use them during training too.
How do we train this? Learning to synthesize a better answer from multiple attempts can be handled with standard RL. The harder problem is teaching models to generate a set of traces that are useful together for a downstream synthesizer to produce a better final response. This leads naturally to a set RL formulation for training models to generate these traces.
LLMs already use many forms of compute at test time, so they should learn to use them during training too.
How do we train this? Learning to synthesize a better answer from multiple attempts can be handled with standard RL. The harder problem is teaching models to generate a set of traces that are useful together for a downstream synthesizer to produce a better final response. This leads naturally to a set RL formulation for training models to generate these traces.
The most capable reasoning systems in AI scale inference compute along several axes: sequential compute to think longer, parallel compute to sample many independent attempts, and aggregative compute to synthesize prior traces into a new improved one. But during training, we only sequential compute. This creates a fundamental mismatch between how we ultimately deploy these systems and how we train them, leaving much of search and synthesis unoptimized.
We introduce SPIRAL, an RL framework for making all inference-compute primitives end-to-end learnable: models learn to coordinate sequential, parallel, and aggregative reasoning using only the reward of the final output. Work with @ifdita_hasan (co-lead), @michaelyli_ , @oshaikh13 , @yoonholeee , @DorsaSadigh , @chelseabfinn , @noahdgoodman 🧵
0
11
71
7,622
Tristan Thrush retweeted
Often in research level math, you can't verify a proof with Lean because the formal libraries needed don't exist yet.
Our fix: Pseudo-Formalization (PF), a proof verification method with the modularity and precision of formal proofs, but the flexibility of natural language.
Our fix: Pseudo-Formalization (PF), a proof verification method with the modularity and precision of formal proofs, but the flexibility of natural language.
4
32
157
46,662
We open-sourced the code for this project!
You can use it to make synthetic LLM training data for any downstream target.
The code also gives you a minimal example for computing data-weight metagradients through LLM training + evaluation.
Check it out: github.com/TristanThrush/…
You can use it to make synthetic LLM training data for any downstream target.
The code also gives you a minimal example for computing data-weight metagradients through LLM training + evaluation.
Check it out: github.com/TristanThrush/…
New paper!
Want to precisely optimize synthetic training data to do practical or even wacky things?
Dataset Policy Gradients get you there, letting you target any differentiable training or post-training metric. We embedded a QR code in GPT-2’s weights using only training data!
1
7
70
13,318
Tristan Thrush retweeted
Fast Byte Latent Transformer is accepted to ICML 2026! ⚡🥪
Byte-level LMs promise to free us from subword tokenizers, but decoding one byte at a time is super slow.
We make BLT generation more efficient with BLT-D: text diffusion for parallel byte decoding. 1/
Byte-level LMs promise to free us from subword tokenizers, but decoding one byte at a time is super slow.
We make BLT generation more efficient with BLT-D: text diffusion for parallel byte decoding. 1/
14
106
743
107,635
Tristan Thrush retweeted
Self-play led to superhuman Go performance, why hasn’t it for LLMs?
In practice, long run self-play plateaus like RL. We study why this happens, and build a self-play algorithm that scales better. It solves as many problems with a 7B model as the pass@4 of a model 100x bigger.
In practice, long run self-play plateaus like RL. We study why this happens, and build a self-play algorithm that scales better. It solves as many problems with a 7B model as the pass@4 of a model 100x bigger.
30
152
1,016
156,669
Tristan Thrush retweeted
I'll be at ICLR in Rio next week! Come catch me presenting the Automated AI Research Trilogy at the Thursday morning poster session!
DM is open if you wanna schedule chats!
DM is open if you wanna schedule chats!
0
19
124
8,888
Tristan Thrush retweeted
Excited about this new work led by @TristanThrush!
Obvious implications for targeted synthetic data generation and data poisoning (check out the original thread for very cool results), but wanted to highlight a few points (1/7)
Obvious implications for targeted synthetic data generation and data poisoning (check out the original thread for very cool results), but wanted to highlight a few points (1/7)
New paper!
Want to precisely optimize synthetic training data to do practical or even wacky things?
Dataset Policy Gradients get you there, letting you target any differentiable training or post-training metric. We embedded a QR code in GPT-2’s weights using only training data!
1
10
53
8,686
Tristan Thrush retweeted
I’m excited about how this work can be used to understand “weird generalization.” AI trained on seemingly benign data can acquire “weird” behaviors that are hard to predict. Examples include emergent misalignment, subliminal learning, and data poisoning with harmless inputs. 🧵
New paper!
Want to precisely optimize synthetic training data to do practical or even wacky things?
Dataset Policy Gradients get you there, letting you target any differentiable training or post-training metric. We embedded a QR code in GPT-2’s weights using only training data!
3
6
58
9,920
Overall, the Dataset Policy Gradient permits finer control of synthetic data than ever before! This has implications for improving performance of LLMs and for data poisoning. See our paper for other experiments and much more detail!
arxiv.org/abs/2604.08423
arxiv.org/abs/2604.08423
1
0
16
1,028
Thanks to the amazing team! @smsampark, @hermanbrunborg, @LukeBailey181, @marcelroed, @neilbband, @ChrisGPotts, @tatsu_hashimoto
0
0
14
887