▲ 86 · yesterday 00:00Making streaming video generation faster with fewer steps
They found a way to create long, high-quality videos using fewer computing steps than before, by removing an extra helper network that the old method needed. This could make fast video generation cheaper and easier to run.
Technical summary+
Few-Step Streaming Video Generation via Condition Shifting
APEX extends Distribution Matching Distillation-style few-step streaming video generation by estimating the generated distribution's velocity field inside the generator via a shifted condition instead of a separate fake-score network, keeping S2-Avatar's frozen teacher, Self-Replay Forcing, and perceptual regularization. On 180-second videos it beats S2-Avatar on Progressive instruction fulfillment (3.856 vs 3.831) and aesthetics (0.6419 vs 0.6326), with +30.0% net human preference on instruction following and 4.4% lower peak memory under matched training.
Helping AI agents double-check their stored lessons
AI agents keep notes from past tasks to help with new ones, but sometimes those notes are wrong or misread, making the agent fail tasks it could have solved alone. This work links each note back to the original story it came from, so the agent can check its sources before acting — like a scientist citing evidence.
Technical summary+
Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents
WikiTrace stores each distilled lesson with links to a success/failure summary and the raw trajectory, and the agent descends those links with terminal tools whenever an insight lacks detail, conflicts with another, or came from a failed task. On five agentic benchmarks it consistently outperforms existing memory approaches and an insights-only baseline.
Fixing AI models by flipping tiny number switches
Instead of retraining a whole AI model, researchers made targeted micro-edits to its internal numbers to fix a flaw where it got stuck in reasoning loops. The edited model did much better on math tests than both the original and a version fixed by an AI agent.
Technical summary+
Tree-Ising: Repairing Language Models with Bit-Level Ising Edits
Tree-Ising casts gradient-free weight repair as sparse discrete optimization over BF16 weights, decomposing a per-row objective into independent QUBO/Ising problems solved via simulated annealing or simulated bifurcation, with edits restricted to sign flips and exponent shifts. On the two edited layers it reaches 54.6% on MATH-500 (bifurcation: 55.1%) vs the agent's 51.7% and base 7.9%, beating the agent edit across five math benchmarks with significant gains on MATH-500 and GSM8K.
Robots learning by practicing in a simulator, not retraining
Instead of locking a robot's skills into its brain through training, this approach lets it rehearse a task in a practice world and write what it learned into a scratchpad it reads later. With just one demonstration per task, the robot succeeded on 50 tasks far more often than the best traditionally-trained system.
Technical summary+
World In-Context Learning for Agentic Robot Manipulation
WICL has a frozen model rehearse a demonstrated task in simulation, writing world feedback into its context rather than weights; RoboScript freezes this as a Task Memory with data-isolated deployment. Across all 50 RoboTwin 2.0 tasks it achieves 83.4% mean success with one demo per task and no gradient updates, exceeding the strongest trained policy by 20+ points.
Testing AI that handles tables on many jobs at once
AI systems that work with data tables can be asked to fill in missing values, make predictions, or invent new realistic data — usually tested separately. This new benchmark checks many systems on all three jobs together, revealing which skills go together and where one model can't do everything well.
Technical summary+
One Table, Many Queries: Benchmarking Generalist Tabular Models Across Generation, Imputation, and Prediction
The One Table Benchmark spans 56 datasets, 46 methods, and 20 capability axes with aligned splits; they train five backbones under all seven non-empty objective combinations and compare generalists vs specialist portfolios. Key findings: biggest pooled gains come from including the recipient task's own objective, generation has the broadest positive auxiliary effects, and cross-objective benefits need not add and vary by backbone.
Using structured textbooks to teach AI math reasoning
AI language models are currently trained on fragmented math problems, which often misses the logical flow of real mathematical arguments. This research organized thousands of university textbooks into a vast dataset to help AI understand not just answers, but the complete structure of a mathematical proof.
Technical summary+
Scaling Mathematical Proof Supervision from Structured Textbooks
TOME-Math converts 2,118 university textbooks into 380K structured proof-supervision records, preserving curricular and dependency signals. Fine-tuning on this corpus increased the unweighted mean score across evaluations by 16.7 percentage points over the base model.
Turning images into language-like codes for AI models
Current AI models often treat pictures and text separately, but this paper explores a way to represent visual information using discrete "tokens"—similar to how words are used in text. This allows AI to use the same unified frameworks that power language models to understand and process images.
Technical summary+
dNaViT: Discrete Codes are General-Purpose Visual Representations
dNaViT is a Discrete Native-resolution Vision Transformer that produces discrete visual primitives analogous to language tokens using grouped product quantization. Evaluations demonstrate that dNaViT matches or exceeds task-specific baselines across 12 visual understanding, 4 generation, and 1 image editing benchmarks.
A fast and flexible method for making large AI models run efficiently
Large AI models use massive amounts of computing power, which can be expensive and slow. This research developed a new compression technique that reduces the memory and processing requirements of these models without significantly hurting their performance.
Technical summary+
MACaQ: Fast, Model Agnostic Calibrated Quantization for Low-Precision Weights and Activations
MACaQ is an MXFP4 W4A4 post-training quantization method that treats every linear layer independently, using online scale and Hadamard transforms. It reached lower perplexity than competitors while matching Llama-3-8B performance in 1/11th the training time.
Training compressed AI agents with a guiding supervisor
When AI models are shrunk to run on less powerful hardware, they often make mistakes or fail in complex tasks. This work introduces a training method where a powerful, full-precision AI acts as a supervisor to correct the errors of the smaller, quantized AI.
Technical summary+
SteerDistill: Quantizing LLM Agents via Closed-Loop Teacher Steering
SteerDistill uses a closed-loop teacher steering mechanism to guide a student model to revise failed quantized responses based on a full-precision teacher. This method exceeded the strongest baseline by 1.67–2.18 points under W4A16 and used 67% fewer training GPU hours.
A comprehensive test for robots controlled by coding AI
While AI can control simulations, it's not clear if an AI that writes code can reliably control a real-world robot. This paper created a massive, unified testing environment to determine if coding agents are ready to operate physical machinery.
Technical summary+
Embodied First Exam
Embodied First Exam is a unified, agent-native evaluation framework consisting of 1,156 tasks across 47 robots, covering everything from pick-and-place to multi-robot interaction. The evaluation uses Standard, Realtime, and Privileged settings to test agent capabilities under different constraints.
The most-upvoted papers on Hugging Face Daily Papers from the last few days. Explanations are written by The Wire from each paper's abstract.