Jefferson Enrique Hernandez Cevallos

Thread on VJEPA 2.1🤟 This DEFINITELY flew under the radar: just a few days ago, @AIatMeta released V-JEPA 2.1, taking a massive step toward closing the gap between image and video domains. For a long time, image backbones were the only option for solving dense vision tasks. This model disagrees, showing that universal spatial understanding also emerges from large-scale video models!🎥

Blog post - Transformers as Constrained Optimization Rewriting pre-norm decoder-only transformers as solutions to regularized objectives. Changing regularization to hard constraint gives a canonical temperature, generalizing to KL-divergence, ideas of cross-layer interaction.

Is RL dead for post-training? Of course not, but there are other interesting options for fine-tuning ▪️ Evolution Strategies (ES) is a gradient-free optimization method that tests random parameter changes and moves the model toward the best-performing ones. - It creates a small population of models by adding random perturbations to the parameters - Perturbed models' outputs are scored with a reward function/verifier - Model parameters are updated in the direction of perturbations that achieved best rewards The best thing is that ES can scale to billion-parameter models and shows clear gains over RL • On Countdown benchmark: ES raised Qwen-2.5-3B to 60.5% (vs 32.5% GRPO) and Llama-3.1-8B to 61.2% (vs ~51% RL) • ARC-AGI 0.2% → 29.5% and Sudoku 2.5% → 69.5% improvement And everything without computing gradients through backpropagation So don’t overlook other approaches in favor of RL alone - there is a lot to explore. Here we’ve gathered the new fine-tuning stack for LLMs with ES and most promising LoRAs -> turingpost.com/p/beyondrl

Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/nca-pre-p… (1/n)

Modern vision models lacks robustness when objects appear in unusual poses. @StphTphsn1 and I study latent equivariant operators as a remedy and discuss caveats of these operators. Below is a summary of the work, accepted at the GRaM Workshop at ICLR @iclr_conf 2026. 🧵

We’re acquiring Promptfoo. We’re buying Safetysnoot. Proud to announce that the Ragschlorp team is joining us. We own a majority stake in Chunkwad.

Most imagination-based world models learn representations by reconstructing pixels. But reconstruction may not be the right objective for control. In our new paper we explore a different idea: 👉 predict the next embedding instead of reconstructing observations. Introducing NE-Dreamer. Project page: corl-team.github.io/nedreamer/ Paper: arxiv.org/pdf/2603.02765 Code: github.com/corl-team/nedr…

Two days ago I didn't know if a 10K transformer could add 10-digit numbers. AdderBoard is now at 36 hand-coded @alexlitzenberge (🏆) !! 311 trained @reza_byt (🏆) !! The gap is v interesting: humans can still construct solutions that gradient descent can't find. For now

Apparently counterintuitive result, but obvious if you think about it: Thinking longer is bad for OCR Just like thinking longer is bad for VQA Too much thinking and the image is too far from the answer, and the answer's tokens attend the reasoning trace, not the image [1/3]

Excited to share our new work EditCtrl! We introduce a disentangled local-global control video inpainting framework that dynamically allocates compute where needed - achieving up to 10x compute savings over full-attention while matching or exceeding SOTA editing quality. 🧵

I'll admit, going in I was not 100% sure this was possible: we trained a tiny 4B model (QED-Nano) to prove math theorems at the Olympiad level! Today, we release the full recipe, from the data curation done for SFT to our RL algorithm that explicitly optimizes for test-time scaling over millions of tokens (i.e., we train QED-Nano to continually improve as we apply modern day test-time scaffolds like DeepSeekMath-agent over it). 🧵⬇️

In my recent blog post, I argue that "vision" is only well-defined as part of perception-action loops, and that the conventional view of computer vision - mapping imagery to intermediate representations (3D, flow, segmentation...) is about to go away. vincentsitzmann.com/blog/bitter_le…

Why is RL training with policy-gradient methods often unstable—especially near the optimum? Our new work studies this through the noise-to-signal ratio (NSR) of the REINFORCE gradient estimator (defined as the estimator variance (noise) normalized by the squared norm of the true gradient (signal)). We show that the NSR is highly non-uniform and often increases as training progresses, which can trigger instability or even policy collapse. Paper link: arxiv.org/abs/2602.01460 Long thread👇

🚨Transformers don't learn Newton's laws? They learn Kepler's laws! Like us, transformers don't predict a flying ball via a differential equation, but by fitting a curve. Moreover, reducing context length steers a transformer from Keplerian to Newtonian. Compression in play.

🚀 Two papers accepted to #ICLR2026 on test-time scaling for vision-language systems (retrieval + reasoning)! 1) MetaEmbed (Oral Presentation): Meta Tokens + Matryoshka multi-vector training → flexible late interaction, choose #vectors at test time for accuracy↔efficiency. Paper: arxiv.org/abs/2509.18095 Work done at @AIatMeta with amazing collaborators: Qi Ma, @Mengting_Gu, Jason Chen, Xintao Chen, @vislang and @MohanVijaimohan! 2) ProxyThinker: training-free test-time guidance from small “slow-thinking” visual reasoners → self-verification / self-correction via distribution-level guidance. Paper: arxiv.org/abs/2505.24872 Work done with @JaywonK17250, @Siru_Ouyang, @jefehern, @yumeng0818 and @vislang! While I won't be able to travel to Brazil🇧🇷, please say Hi to the team :-) #MultimodalRetrieval #VisualReasoning #VisionLanguage #TestTimeCompute #Embeddings

Rewards as Labels: Revisiting RLVR from a Classification Perspective "we propose Rewards as Labels (REAL), a novel framework that revisits verifiable rewards as categorical labels rather than scalar weights, thereby reformulating policy optimization as a classification problem. "

And RL on imagenet is such a nice way to show this: every task can be formulated as RL but should you? Here they show how Reinforce has terrible efficiency but this MaxRL thingy with *large number of rollouts* don't.