Articles & Guides

LiveAnimate targets a familiar tradeoff in pose-driven human animation: diffusion models can look good, but they are too slow and too memory-hungry for interactive use. The paper reframes a 14B video DiT into a streaming system built for stability over long sequences.

UniSwap proposes a unified framework for replacing both the visual identity and vocal timbre in talking videos. Instead of stitching together separate face and voice models, it performs joint audio-visual transfer inside a single diffusion transformer.

StateFlow proposes a state-centric framework for generative previsualization across film, games, architecture, and urban design. Instead of producing a video in one shot, it maintains an editable 3D world that can be constructed, evolved, and accessed over time.

A new paper turns “staying coherent over many turns” into an automatic benchmark. NCP-Bench shows that even strong models can drift, contradict themselves, and lose track of key story commitments.

UniMoMo is a post-training compression method for large recommendation MoE models. It merges experts by functional behavior on calibration data and protects high-traffic routing paths to preserve quality while improving serving speed.

VibeLifeBench reframes agent evaluation around long-running everyday tasks where the world changes silently and constraints are often implicit.

Ex-Omni-2D tries to close a gap in today’s omni-modal assistants: they can talk, but they still feel visually absent. The paper combines text, personalized speech, and reference-conditioned video into one dialogue pipeline.

Evidence-RL targets a persistent weakness in vision-language models: correct answers may come from language priors or dataset shortcuts rather than image evidence. Its CED mechanism audits whether a response causally depends on the local visual region that should support it.

SPOT improves on-policy distillation by deciding where to spend extra probing budget and what targets the student should learn. Instead of relying only on local teacher probabilities, it calibrates distillation with verifier-scored downstream outcomes.

Agent Memory Distillation reframes knowledge transfer for tool-using agents: instead of retraining, it distills successful teacher trajectories into layered memories that small models can reuse. The goal is to make weak agents act more reliably without collecting massive new data.

A new technical report argues that interpretability does not have to be bolted onto language models after training. Instead, it can be built into the training process and optimized alongside language modeling itself.

A new study argues that follow-up suggestions in image-creation chats should be visually grounded, not merely text-driven. Using real product data, the proposed three-stage framework improves suggestion relevance, user engagement, and visual consistency.

Motif 3 is a decoder-only sparse MoE language model with 314B total parameters and 13.2B activated per token. Its design combines GDLA, training-stability tricks, and multi-teacher post-training to improve reasoning, coding, and long-context performance.

The paper introduces U-OPSD, an unsupervised on-policy self-distillation method that builds pseudo-solutions from a model’s own multiple generations. By training on disagreements with those self-consistent pseudo-solutions, it improves mathematical reasoning without ground-truth labels or stronger teachers.

Ouroboros reframes coding agents as systems that can improve not only outputs, but also their tools, prompts, context assembly, and core implementation. The paper reports leading results on Terminal-Bench 2.1, OSWorld-Verified, and CL-Bench while emphasizing governed evolution.

SWE-Bench ProMax shifts coding-agent evaluation from narrow bug fixing to large-scale, behavior-preserving refactoring across multiple files and languages.

YOLO-PEFT reframes PEFT adapter placement for real-time detectors as a constraint-planning task. Instead of relying on manual trial and error, it decides where adapters can be safely inserted—or refuses the request before training.

The Skaling law introduces a single interaction exponent to couple model capacity and training data, addressing systematic errors in conventional neural scaling laws.

This paper turns Test-Time Training into a modular design space instead of a set of hard-coded variants. Its main contribution is not only a framework, but also a clearer way to isolate which TTT components actually matter.

Despite recent losses and a major reduction in its public portfolio, AI-focused hedge fund Situational Awareness is still making large private bets. Its latest move is a $400 million investment in chip-manufacturing startup Source Foundry.

Us vs. Them is an open-source tool that uses version history and diffs to infer whether ranges of text were mainly written by humans or agents. It targets a growing problem in agentic editing: knowing which parts of a file should be treated with extra care.