Abhijeet Awasthi

Agentic-RL and Code-LLMs @AmazonScience/@kirodotdev
Prev: @Meta, @MSFTResearch, @iitbombay, KGP

[Arash Bidgoli @ArashBidgoli]

people complain all the time about 996. wait till they hear about 007

[Abhijeet Awasthi @awasthi_a_]

semantic parsing ftw.

[Omar Khattab @lateinteraction]

@YouJiacheng wow what? self-training LLMs to use tools and especially search is much older in the literature than 2021.. :-)

[Mahesh Sathiamoorthy @madiator]

Perplexity uses Nimble codebase in various ways, and no credit given.
@AravSrinivas please fix :)

[Abhijeet Awasthi @awasthi_a_]

Mixture of MOEs done right

[Saurabh Shah @saurabh_shah2]

nvidia acquiring reflection makes sense
Nemo is a 550B
Beam is 501B

550B+ 501B = 1T so they’ll have a K 2.6 level model

If they can get thinky on board that adds another ~1.2T (inkling big and small) so the next nemotron could be as many as 2.2T, in striking range of K3

(ty @srush_nlp for teaching me how to do back of the envelope calcs for transformers)

[Andrew Drozdov @mrdrozdov]

Listening to an @lateinteraction keynote is better than any cup of coffee.

One of the points that resonates strongly w/ me:
✅ Stronger models will need longer prompts rather than shorter, since we'll be more demanding.

[Rohan Pandey @khoomeik]

PSA to data cos: you need to build frontier-level training infra if you want to prove the value of your training data

the data game is past its slop era

[Harman Singh @ ICML 🇰🇷🇰🇷 @Harman26Singh]

Persistent self-improvement is model- and task-dependent. Do the gains generalize? Checkpoints are scored on held-out games.

In Hard chess, Claude Fable 5 climbs from 37.5% to 73.3% on them late in training; Claude Opus 5 from 25.0% to 66.4%. Against a stronger engine, gains are generally smaller.

This motivates evaluating plasticity across a distribution of agentic tasks during frontier agent evals.

[Abhijeet Awasthi @awasthi_a_]

@amuseddaman @g_k_swamy @Zanette_ai Congratulations!!! 🎉

[Liam Fedus @LiamFedus]

@DBahdanau is 🐐

Over the years, I've had the awesome fortune of working with the best ML researchers in the field, plus the incredible new cohort of future leaders who IC the AI tech we use daily.

And Dima still stands out as a rare trifecta: great leader x brilliant researcher x exceptional systems engineer.

[Abhijeet Awasthi @awasthi_a_]

@satnam6502 love the vibes at Fremont's Mylapore and their Mysore Masala Dosa of course

[Kamal Gupta @kamalgupta09]

good things come to those who look at data

[Mahesh Sathiamoorthy @madiator]

Introducing Bespoke Nimble: an open data, open model, open recipe for an open Jev.

Code and info: https://t.co/aC7kPejrcj
Model: https://t.co/snwKGdhn1I

Data:
* A new data curation recipe called contrastive data curation.
* Slightly change facts to generate negative data. This pushes the model to discriminate better and become a better decision maker. The calibration is implicit.
* Didn't do ablations but I think this is a critical piece!
* This also means training data doesn't need probabilities.
* Data covered 10 categories, and is fully synthetic.
* This data is split into train and eval.

Training
* LoRA finetune of Qwen3.5-9B.
* Distillation-free: we use Jev to only evaluate.
* No RL yet!

Serving
* Parallel constrained decoding as suggested by @NielsRogge and @harshagundal.

Results:
* The post-trained Qwen (Nimble) became substantially better on our curated eval: 66% for Qwen to 90% for Nimble. Jev is at 93%.
* 100ms on H100 and free to use on your macbook! Feel the AGI for free.
* 2 days of building in public. :)

Big caveat is that there is no standard benchmark to measure performance, and it's possible Nimble is much worse on other benchmarks compared to Jev. But it should be better than Qwen!

We thank @typesafeai for making Jev and the inspiring discussions in the community. Hope this release lifts all the boats and encourages more research and activity in this space.

[Abhijeet Awasthi @awasthi_a_]

@mr_bytestorm You should baseline your expectations w.r.t. the base model (q38-27B) being optimized here :-)

[Martin Marek @mrtnm]

Why is RL so sensitive to training-inference mismatch (TIM)?

With each training step, the trainer is pulled toward the sampler, resulting in accumulating drift.

We exploit this intuition to devise a novel correction method to stabilize RL under TIM, called Score Centering.

[Max Ryabinin @m_ryabinin]

RL with LLMs is very unstable when training and sampling policies differ. Standard fixes (matching numerics, importance sampling) work around the problem.

We find the root cause of this instability from first principles and propose a way to directly cancel it. Score Centering is competitive and compatible with existing approaches — while simple to implement!

🧵 [1/6]

[Abhijeet Awasthi @awasthi_a_]

@_LuoFuli 👀

[DeepSeek @deepseek_ai]

🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.

🔹 Introducing the smallest model in our new architecture family, with native visual understanding.
🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.

1/6

[Ofir Press @OfirPress]

interesting word choice in the astra report

[Sunny Sanyal @SunnySanyal9]

A few non-obvious limitations of loop transformers (pure depth recurrence).

Looping ≠ equivalent depth.

For the same Llama model (w/ Fineweb), looping 2× underperforms simply doubling the depth. The deeper model is clearly better.

[kalomaze @kalomaze]

> post-training is best understood as redistributing probability mass inside the pretrained distribution

ok but to the extent that this is true it seems... vacuous??? yes, and this is also true of how pretraining adapts the random distribution you get at initialization