Posterior Augmented Flow Matching
Paper • 2605.00825 • Published
AI & ML interests
Aligning LLMs to be helpful, honest, harmless, and huggy (H4)
Recent Activity
View all activity
sayakpaul
authored 3
papers 4 days ago
sayakpaul
submitted a
paper to Daily Papers 4 days ago
sergiopaniego
posted an update 5 days ago
Post
7520
Frontier models use distillation as a step of their post-training pipelines.
In 2026 it has three jobs: compress a big model into a small one, merge RL experts into a single model, and let a model teach itself.
I wrote up which frontier models use each one and how: https://huggingface.co/blog/sergiopaniego/distillation-2026
It pairs with Class 2 of the Training an Agent series Ben and I are doing, where we teach these techniques hands-on with TRL!
In 2026 it has three jobs: compress a big model into a small one, merge RL experts into a single model, and let a model teach itself.
I wrote up which frontier models use each one and how: https://huggingface.co/blog/sergiopaniego/distillation-2026
It pairs with Class 2 of the Training an Agent series Ben and I are doing, where we teach these techniques hands-on with TRL!
- 3 replies
albertvillanova
posted an update 8 days ago
Post
3425
🎉 KTO is now part of the stable TRL API
As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175
This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.
Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.
Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.
KTO now sits on equal footing with our other flagship trainers. 🚀
As of Promote KTO to stable API, KTOTrainer and KTOConfig have graduated from trl.experimental to the stable trl API. https://github.com/huggingface/trl/pull/6175
This one closes out a long road. Over the past 6+ months, the "Align KTO with DPO" effort landed ~90 PRs methodically bringing KTO up to the standard we hold for stable trainers, one carefully-scoped change at a time:
- Feature parity with DPO: full VLM support (incl. multi-image), sync_ref_model, PEFT + Liger, ZeRO-3 + PEFT dtype fix, pad_to_multiple_of, activation offloading, IterableDataset and dict eval_dataset, remove_unused_columns, and reference-logprob precomputation at init.
- Consistency with DPO: aligned method order and signatures, tokenization, _prepare_dataset, PEFT handling, ref-model preparation for distributed training, and config layout — plus a new DataCollatorForKTO and output format. Metrics moved into _compute_loss and simplified to direct averages via the shared _metrics attribute.
- Removing legacy baggage: dropped encoder-decoder support, BOS/EOS handling, null_ref_context, generate_during_eval, model_init, preprocess_logits_for_metrics, model/ref adapter names, and several dead config knobs.
- Coverage: a full test suite mirroring DPO, text collator tests, VLM tests, and slow tests.
- The promotion itself: the experimental → stable move (#6175) and shim cleanup (#6287), handled so downstream users get a clean deprecation path.
Honestly, this has been one of the more complex tasks I've taken on since joining the team, not because any single change was hard, but because it demanded sustained consistency across a ~2,000-line trainer, with every branch, comment, and edge case kept in lockstep with DPO.
Huge thanks to everyone who reviewed along the way (especially @qgallouedec ), the incremental review cadence is exactly what kept this maintainable.
KTO now sits on equal footing with our other flagship trainers. 🚀
- 2 replies
Post
399
Uhh did Opus 4.8 cheat on PostTrainBench??
it found an API key in the PostTrainBench environment that allowed it to generate synthetic training data without using GPU hours, boosting the base model by 0.4913
Source: https://posttrainbench.com/traces/run.html?id=claude_non_api_max_claude-opus-4-8_10h_run1__healthbench_Qwen_Qwen3-4B-Base_17315102#tab=trace
it found an API key in the PostTrainBench environment that allowed it to generate synthetic training data without using GPU hours, boosting the base model by 0.4913
Source: https://posttrainbench.com/traces/run.html?id=claude_non_api_max_claude-opus-4-8_10h_run1__healthbench_Qwen_Qwen3-4B-Base_17315102#tab=trace
- 1 reply
sergiopaniego
posted an update 17 days ago
Post
308
TRL v1.7.0 is out‼️
+ continuous batching makes GRPO and RLOO 1.25x faster at -16 GB
+ proper MoE post-training across GRPO/RLOO/AsyncGRPO
+ new GMPO trainer
+ AsyncGRPO weight sync + padding-free
+ more
https://github.com/huggingface/trl/releases/tag/v1.7.0
wrote a small article about the continuous batching for GRPO feature
https://huggingface.co/blog/sergiopaniego/cb-trl-grpo
+ continuous batching makes GRPO and RLOO 1.25x faster at -16 GB
+ proper MoE post-training across GRPO/RLOO/AsyncGRPO
+ new GMPO trainer
+ AsyncGRPO weight sync + padding-free
+ more
https://github.com/huggingface/trl/releases/tag/v1.7.0
wrote a small article about the continuous batching for GRPO feature
https://huggingface.co/blog/sergiopaniego/cb-trl-grpo
sergiopaniego
posted an update 24 days ago
Post
310
Continuous batching just landed in TRL for GRPO!
At 64 generations it runs faster and uses less VRAM than plain generate, no vLLM needed
How it works and when to reach for it, below
https://huggingface.co/blog/sergiopaniego/cb-trl-grpo
At 64 generations it runs faster and uses less VRAM than plain generate, no vLLM needed
How it works and when to reach for it, below
https://huggingface.co/blog/sergiopaniego/cb-trl-grpo
sergiopaniego
posted an update 27 days ago
Post
293
GLM-5.2 is open and comes with competitive performance against opus 4.8
day-0 in transformers + vllm + sglang, mit license 🤗
on the post-training side: critic-based ppo for variable-length agentic rollouts (ppo is back!) + an online anti-reward-hacking module that feeds the agent dummy info when it tries to cheat
day-0 in transformers + vllm + sglang, mit license 🤗
on the post-training side: critic-based ppo for variable-length agentic rollouts (ppo is back!) + an online anti-reward-hacking module that feeds the agent dummy info when it tries to cheat
sergiopaniego
posted an update about 1 month ago
Post
3928
OpenEnv has a new home: github.com/huggingface/OpenEnv
Starting today, it's coordinated by a committee that includes Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, and Hugging Face
frontier labs train their models and their harnesses together. Claude knows Claude Code. GPT-5.5 knows Codex. that's not an accident, it's training. open-source models deserve the same magic, but pulling that off requires infrastructure that belongs to everyone, not one lab
OpenEnv is that layer. one api, any harness, any trainer, any environment
Rewards and training loops stay in TRL, Unsloth, wherever you already work. OpenEnv is the socket they all plug into
Get involved!
Full announcement: https://huggingface.co/blog/openenv-agentic-rl
Starting today, it's coordinated by a committee that includes Meta-PyTorch, Reflection, Unsloth, Modal, Prime Intellect, Nvidia, Mercor, Fleet AI, and Hugging Face
frontier labs train their models and their harnesses together. Claude knows Claude Code. GPT-5.5 knows Codex. that's not an accident, it's training. open-source models deserve the same magic, but pulling that off requires infrastructure that belongs to everyone, not one lab
OpenEnv is that layer. one api, any harness, any trainer, any environment
Rewards and training loops stay in TRL, Unsloth, wherever you already work. OpenEnv is the socket they all plug into
Get involved!
Full announcement: https://huggingface.co/blog/openenv-agentic-rl
sergiopaniego
posted an update about 1 month ago
Post
327
Frontier agents are this good partly because the model was trained inside the very harness it ships with.
NVIDIA's new paper "Polar: Agentic RL on Any Harness at Scale" brings that recipe to the open: it turns coding harnesses like Codex, Claude Code, Qwen Code or Pi into RL training environments without touching their internals.
The core idea: every agent, however complex or closed, talks to a model through an API, so they put a proxy there. The harness runs exactly like in production while the proxy records prompts, sampled token ids and logprobs. Trajectories get rebuilt outside, token faithful, so gradients hit the exact tokens the policy sampled.
The gains are consistent across all four harnesses. Same Qwen3.5-4B, plain GRPO, evaluated on SWE-Bench Verified:
Codex 3.8 → 26.4 (+22.6)
Claude Code 29.8 → 34.6 (+4.8)
Qwen Code 34.6 → 35.2 (+0.6)
Pi 34.2 → 40.4 (+6.2)
The biggest gains appear on unfamiliar execution paths, Codex being the clearest case. The takeaway: you are not just training a model, you are training the model + harness system.
Two engineering pieces make it work at scale. Async worker pools isolate container boots (CPU), agent execution (GPU) and long tail test runs, so slow runtimes never block the GPUs. And prefix merging stitches hundreds of captured API calls back into contiguous traces: 5.4x faster trainer updates and rollout GPUs at 88% utilization.
It also doubles as an SFT data factory: 504 test verified agent traces from a 122B teacher, multi-turn conversations averaging 104 messages each, coming to the Hub under Apache 2.0 (release pending review).
Paper authors: Binfeng Xu, Hao Zhang, Shaokun Zhang, Songyang Han, Mingjie Liu, Jian Hu, Shizhe Diao, Zhenghui Jin, Yunheng Zou, Michael Demoret, Jan Kautz and Yi Dong.
> Paper: Polar: Agentic RL on Any Harness at Scale (2605.24220)
> Code: https://github.com/NVIDIA-NeMo/ProRL-Agent-Server
> Training data: NovaSky-AI/SkyRL-v0-293-data
NVIDIA's new paper "Polar: Agentic RL on Any Harness at Scale" brings that recipe to the open: it turns coding harnesses like Codex, Claude Code, Qwen Code or Pi into RL training environments without touching their internals.
The core idea: every agent, however complex or closed, talks to a model through an API, so they put a proxy there. The harness runs exactly like in production while the proxy records prompts, sampled token ids and logprobs. Trajectories get rebuilt outside, token faithful, so gradients hit the exact tokens the policy sampled.
The gains are consistent across all four harnesses. Same Qwen3.5-4B, plain GRPO, evaluated on SWE-Bench Verified:
Codex 3.8 → 26.4 (+22.6)
Claude Code 29.8 → 34.6 (+4.8)
Qwen Code 34.6 → 35.2 (+0.6)
Pi 34.2 → 40.4 (+6.2)
The biggest gains appear on unfamiliar execution paths, Codex being the clearest case. The takeaway: you are not just training a model, you are training the model + harness system.
Two engineering pieces make it work at scale. Async worker pools isolate container boots (CPU), agent execution (GPU) and long tail test runs, so slow runtimes never block the GPUs. And prefix merging stitches hundreds of captured API calls back into contiguous traces: 5.4x faster trainer updates and rollout GPUs at 88% utilization.
It also doubles as an SFT data factory: 504 test verified agent traces from a 122B teacher, multi-turn conversations averaging 104 messages each, coming to the Hub under Apache 2.0 (release pending review).
Paper authors: Binfeng Xu, Hao Zhang, Shaokun Zhang, Songyang Han, Mingjie Liu, Jian Hu, Shizhe Diao, Zhenghui Jin, Yunheng Zou, Michael Demoret, Jan Kautz and Yi Dong.
> Paper: Polar: Agentic RL on Any Harness at Scale (2605.24220)
> Code: https://github.com/NVIDIA-NeMo/ProRL-Agent-Server
> Training data: NovaSky-AI/SkyRL-v0-293-data
sergiopaniego
posted an update about 1 month ago
Post
280
The recording from our talk: "From Responses To Trajectories: Multi-Turn and Multi-Environment RL" from PyTorch Conf Europe is live!
@kashif and I covered the latest advances in multi-turn GRPO in TRL: trajectories, tool use, envs, and agentic post-training at scale
https://www.youtube.com/watch?v=rPBeXFntJSU
@kashif and I covered the latest advances in multi-turn GRPO in TRL: trajectories, tool use, envs, and agentic post-training at scale
https://www.youtube.com/watch?v=rPBeXFntJSU
sergiopaniego
posted an update about 1 month ago
Post
217
how do you sync a trillion parameter model every RL step without a shared cluster? we just wrote a blog about it, led by @aminediroHF
what I like the most is the way it proves you can use the Hub for basically everything 🧐 → trainer on one machine, vLLM in a HF Space, the wordle env in another HF Space and weights going through a Hub Bucket. no shared cluster, just HTTPS
it works because ~99% of bf16 weights don't change between RL steps so you only sync the diff. 1.2 GB to 25 MB of payload per step
https://huggingface.co/blog/delta-weight-sync
what I like the most is the way it proves you can use the Hub for basically everything 🧐 → trainer on one machine, vLLM in a HF Space, the wordle env in another HF Space and weights going through a Hub Bucket. no shared cluster, just HTTPS
it works because ~99% of bf16 weights don't change between RL steps so you only sync the diff. 1.2 GB to 25 MB of payload per step
https://huggingface.co/blog/delta-weight-sync
sergiopaniego
posted an update about 1 month ago
Post
2353
most multi-turn RL loops have a silent bug: you decode the model's output to detect tool calls, then re-tokenize the conversation for the next turn. BPE isn't invertible, so decode then re-encode can land on different ids. gradient ends up on tokens the model never sampled. no crash, just quietly wrong math and broken training
@qgallouedec wrote a super educational blog on MITO (message-in, token-out) vs TITO (token-in, token-out) and how you might fix the problem above
go read it 🤓
https://qgallouedec-tito.hf.space/
@qgallouedec wrote a super educational blog on MITO (message-in, token-out) vs TITO (token-in, token-out) and how you might fix the problem above
go read it 🤓
https://qgallouedec-tito.hf.space/
sergiopaniego
posted an update about 1 month ago
Post
6295
new banger blog alert 🚨
@ariG23498 is starting a blog series about profiling in pytorch and part 1 just dropped
takes you from the simplest scenario to actually knowing what your gpu is doing. if you have never opened a profiler trace this is where you start
covers
find it here: https://huggingface.co/blog/torch-profiler
@ariG23498 is starting a blog series about profiling in pytorch and part 1 just dropped
takes you from the simplest scenario to actually knowing what your gpu is doing. if you have never opened a profiler trace this is where you start
covers
torch.profiler from scratch. reading tables and traces, overhead bound vs compute bound, the full dispatch chain from python to gpu kernels, and what torch.compile is actually fusing under the hoodfind it here: https://huggingface.co/blog/torch-profiler
- 1 reply
sergiopaniego
posted an update about 2 months ago
Post
210
If you have a github repo, you basically have an RL training environment
We're introducing Repo2RLEnv (built by @AdithyaSK ), a tool that mines PRs, commits, CVEs and turns them into verifiable sandboxed tasks with real reward signals, automatically
Outputs to Harbor spec so you can plug it straight into RL training or coding-agent eval
> repo: https://github.com/huggingface/Repo2RLEnv
> collection with envs: https://huggingface.co/collections/AdithyaSK/repo2rlenv-verifiable-rl-environments
We're introducing Repo2RLEnv (built by @AdithyaSK ), a tool that mines PRs, commits, CVEs and turns them into verifiable sandboxed tasks with real reward signals, automatically
Outputs to Harbor spec so you can plug it straight into RL training or coding-agent eval
> repo: https://github.com/huggingface/Repo2RLEnv
> collection with envs: https://huggingface.co/collections/AdithyaSK/repo2rlenv-verifiable-rl-environments
sergiopaniego
posted an update about 2 months ago
Post
266
periodic reminder 🧐
some HF blog posts use a special template, long-form, animated, super deep
I keep them all in one collection that gets updated every time a new one drops so you don't lose track
https://hf.co/collections/sergiopaniego/research-and-long-form-blog-posts
spot one missing? let me know
some HF blog posts use a special template, long-form, animated, super deep
I keep them all in one collection that gets updated every time a new one drops so you don't lose track
https://hf.co/collections/sergiopaniego/research-and-long-form-blog-posts
spot one missing? let me know
sergiopaniego
posted an update about 2 months ago
Post
10021
Harness, Scaffold, Context Engineering, Agent... do you actually know what they mean?
We wrote an AI agent glossary and tried to make sense of it all with simple definitions and real examples
↓ go read it ↓
https://huggingface.co/blog/agent-glossary
We wrote an AI agent glossary and tried to make sense of it all with simple definitions and real examples
↓ go read it ↓
https://huggingface.co/blog/agent-glossary
- 2 replies
alvarobartt
posted an update about 2 months ago
Post
461
Open agents on AWS SageMaker AI with open models from the Hugging Face Hub!
> Deploy an open model from the Hugging Face Hub on SageMaker AI
> Connect the deployed model to Strands Agents
> Add built-in and custom tools for tool calling
> Expose external capabilities through MCP integration
> Bonus: talk to your agent and visualize traces with Gradio
https://alvarobartt.com/agents-on-aws-sagemaker
> Deploy an open model from the Hugging Face Hub on SageMaker AI
> Connect the deployed model to Strands Agents
> Add built-in and custom tools for tool calling
> Expose external capabilities through MCP integration
> Bonus: talk to your agent and visualize traces with Gradio
https://alvarobartt.com/agents-on-aws-sagemaker