Matching stories Newest first
Hugging Face Papers lists SolarRenderer, a few-step structure-conditioned video renderer
SolarRenderer adapts a frozen MiniMax-H3 33B backbone with two LoRAs and renders in four steps, 12.3 times faster than 50-step sampling. A block-causal variant renders 768p video in real time.
Focused Omni-Modal GUI Agents paper proposes FocusGUI adapter for GUI agents
FocusGUI is a lightweight adapter that prioritizes multimodal inputs per step and injects them into a frozen decoder, trained on action supervision alone. Reported gains include +33.2 Type Match and 3.0x Exact Match over the frozen backbone across four omni-modal backbones.
Sage-A1.5, a 35B mixture-of-experts agentic model, is introduced in a research report
Sage-A1.5 builds on A1 with subagent orchestration, progress and evidence verification, and open scientific discovery capabilities. The authors say it is meant as an accessible foundation for the community.
AdaExploreBench benchmarks LLM agents on proactive exploration and adaptation to rule drift
The benchmark hides environment rules and tests whether agents can explore to find them and detect when they change. 9% calibrated accuracy, and supplying sufficient evidence or removing rule drift lifts average performance by close to 10%.
MemPilot proposes on-demand multimodal memory curation for LLM agents
MemPilot uses reinforcement learning to pick between query-agnostic memory retrieval and query-specific curation by LLMs and VLMs, trading off performance, cost and latency across five multimodal agent-memory benchmarks.
SpecMRL redesigns Matryoshka representation learning with a parametric spectral basis
SpecMRL orders features by importance in a learned spectral space, outperforming standard MRL at all reduced dimensions on MTEB. It surpasses full-dimensional MRL using only half the embedding dimensions.
Kronecker Delta Rule proposed for stable long-context linear attention in KronLSTM
KronLSTM extends the xLSTM family with a memory-revision update that changes memory along one direction, and the authors claim bounded memory norm without extra normalization. Code and models are open-source, and the paper reports competitiveness with (Gated) DeltaNet models.
OR-License benchmark tests LLM agents on real operations research engagements
Across 11 frontier LLMs in coding-agent harnesses, several beat the expert reference on objective quality, but none matches its Deployment-Adjusted Quality score. 87 against the expert reference of 1.
Meta publishes MemT-1.0, a memory-native model family with Nano and Flash variants
MemT-1.0 internalizes memory into model computation using Gated DeltaNet-2 and Mixture-of-Storage, trained on 34B tokens. At 128K history, Nano reports about 1.86x inference speedup and 66x lower persistent storage than a full KV cache.
RoAR-VAE improves visual tokenizer latents for diffusion through robust prediction
The paper proposes RoAR-VAE, which adds semantic prediction and reconstruction from corrupted latents. 1-VAE.
Self-HPO framework alternates harness and policy optimization for coding agents
Self-HPO lets a coding agent propose its own harness candidates and then update its policy with reinforcement learning, raising Qwen3.5-35B-A3B's average resolve rate from 57.8% to 64.8% across three coding benchmarks. The paper also reports gains with a stronger external model, such as Opus 4.8, a…
MT Forcing proposes a training method for stable multi-turn image editing
Researchers introduce MT Forcing, which combines Single-Turn Distribution Matching Distillation and Teacher-Forcing Guidance to reduce drift in multi-turn editing with unified multimodal models.
Hugging Face paper introduces CoworkEnv for synthesizing training environments for AI coworker agents
The framework generates executable training environments for AI coworkers and trains PocketWorker models. 9-point gain on multi-turn tasks.
Hugging Face paper presents EgoXperience for world-grounded egocentric reconstruction
The paper introduces EgoX-Data-6M, a 50-hour dataset, and EgoX, a model that jointly reconstructs camera, scene, hands and objects from monocular video in one forward pass. 6% lower hand error against the strongest evaluated baselines.
Interlock framework synthesizes long-horizon agent tasks that improve agent pass rates after training
Interlock represents tasks as completion graphs and generates workspaces, requests and rubrics. 16 points.
OpenVLHarness adds persistent multimodal memory and tool use to vision-language models
The harness gives multimodal models a tool capability layer and persistent memory for images, data and web references. 6 point average gains over no-tool baselines across 22 benchmarks and eight backbones, including GPT-6 Luna at lower token cost.
InfiniDream proposes open-ended procedural generation of photorealistic 3D worlds
The framework pairs a VLM with existing procedural algorithms to create new photorealistic asset types and build large 3D scenes from text. Evaluations cover 100 object and 20 scene generation tasks, with gains on four perceptual metrics.
Hugging Face paper proposes collective experience learning for fleets of robots
FERT and FATS let deployed robots share execution experience to improve a common policy without extra expert corrections. 3%.
Brain encoding method BrainLoRA sets Algonauts 2025 record using parameter-efficient fine-tuning
The paper adapts backbones up to 104B parameters with LoRA and reports the best published result on the Algonauts 2025 Challenge using Qwen3-Omni. An ensemble of 200+ adapted models ranks first on the public post-challenge leaderboard.
Paper introduces AnyGoal benchmark testing world model generalization beyond expert rollouts
AnyGoal is a benchmark of goal-shifted tasks built from common world-model environments, with an open-source library. The authors report that current world models consistently fail when familiar state and action factors are recombined.
Hugging Face paper introduces HarnessPro for cheaper, stronger AI agent harnesses
HarnessPro evolves agent harnesses without ground-truth tests and outputs lite, medium and heavy variants. 1% versus RHO.
Meta publishes ScaleNav, a video-pretrained model for language-guided robot navigation
ScaleNav learns navigation dynamics from internet egocentric videos, then uses simulated action data for training. 9% on HabitatGS-LN, and runs on a real quadruped robot.