AI Paper Summary

Can an iterative draft–revise solver that repeatedly updates a latent scratchpad outperform far bigger autoregressive LLMs on ARC-AGI? Samsung SAIT (Montreal) has launched Tiny Recursive Model (TRM)—a two-layer, ~7M-parameter recursive reasoner that studies 44.6–45% check accuracy on ARC-AGI-1 and 7.8–8% on ARC-AGI-2, surpassing outcomes reported for considerably bigger language fashions reminiscent of DeepSeek-R1, o3-mini-high, and…

TL;DR: A brand new analysis from Apple, formalizes what “mid-training” ought to do earlier than reinforcement studying RL post-training and introduces RA3 (Reasoning as Action Abstractions)—an EM-style process that learns temporally constant latent actions from knowledgeable traces, then fine-tunes on these bootstrapped traces. It exhibits mid-training ought to (1) prune to a compact near-optimal motion…

How a lot compression ratio and throughput would you recuperate by coaching a format-aware graph compressor and delivery solely a self-describing graph to a common decoder? Meta AI launched OpenZL, an open-source framework that builds specialised, format-aware compressors from high-level knowledge descriptions and emits a self-describing wire format that a common decoder can learn—decoupling compressor…

Salesforce AI Research launched CoDA-1.7B, a diffusion-based language mannequin for code that generates by denoising entire sequences with bidirectional context, updating a number of tokens in parallel moderately than left-to-right next-token prediction. The analysis group revealed each Base and Instruct checkpoints and an end-to-end coaching/analysis/serving stack. Understanding the structure and coaching CoDA adapts a 1.7B-parameter…

Researchers from Cornell and Google introduce a unified Regression Language Model (RLM) that predicts numeric outcomes straight from code strings—protecting GPU kernel latency, program reminiscence utilization, and even neural community accuracy and latency—with out hand-engineered options. A 300M-parameter encoder–decoder initialized from T5-Gemma achieves sturdy rank correlations throughout heterogeneous duties and languages, utilizing a single text-to-number…

Table of contents FP8 index → top-k selection → sparse core attention Lets Talk about it’s efficiency and accuracy Summary FAQs DeepSeek launched DeepSeek-V3.2-Exp, an “intermediate” replace to V3.1 that provides DeepSeek Sparse Attention (DSA)—a trainable sparsification path geared toward long-context effectivity. DeepSeek additionally decreased API costs by 50%+, constant with the acknowledged effectivity beneficial…

Can a single AI stack plan like a researcher, cause over scenes, and switch motions throughout totally different robots—with out retraining from scratch? Google DeepMind’s Gemini Robotics 1.5 says sure, by splitting embodied intelligence into two fashions: Gemini Robotics-ER 1.5 for high-level embodied reasoning (spatial understanding, planning, progress/success estimation, tool-use) and Gemini Robotics 1.5 for…

Meta FAIR launched Code World Model (CWM), a 32-billion-parameter dense decoder-only LLM that injects world modeling into code era by coaching on execution traces and long-horizon agent–atmosphere interactions—not simply static supply textual content. What’s new: studying code by predicting execution? CWM mid-trains on two giant households of remark–motion trajectories: (1) Python interpreter traces that document…