Papers

Route Anything presents a unified theory and omni-modal trajectory matrix for routing among models, tools, and policies, finding a 5.4-point per-input oracle gap, modest learned-router gains, and 8.82× average efficiency from cheaper routes.

Pandora simulates psychotherapy clients by separating psychological state and content planning from an individual expression planner, improving expression similarity and reducing perceived AI-like judgments in the reported evaluation.

LoCo learns state-conditioned continue-or-conclude preferences from self-generated math traces by freezing overthinking prefixes and training DPO on shorter verified continuations, reducing reasoning length while preserving or improving accuracy.

MoviEdit is an end-to-end framework for long-form multi-shot video editing that combines scalable paired-data generation, a cascaded bidirectional–autoregressive editor, and a dedicated benchmark.

The paper reports Kling's experience that video generation quality is the projection of multimodal understanding across the data, conditioning, alignment, and evaluation chain.

DreamX repurposes a pretrained video diffusion model to extract compact latent actions that reconstruct future video and transfer across human and robot embodiments.

RoboColosseum is a distributed in-the-wild robot benchmark that evaluates VLA policies and frontier agents across four embodiments via Open and Fine-tune tracks, revealing weak cross-embodiment ranking agreement and sample-efficient active matchmaking.

FactEmbed decouples composed-query comprehension from embedding learning by training factorized MLLM latents with target imagination, factor-level matching, and attention distillation, then reranking with mean-pooled latents.

Axion 1.0 is a fully open agent foundation model trained with two SFT stages plus RL, without dedicated mid-training, DPO, or OPD, and is competitive across a broad agentic benchmark suite.

ToolSpace reverses data-centric environment synthesis by mining tool-composition graphs from real execution traces and synthesizing replay-verified cross-environment tasks, lifting long-horizon tool-use scores and producing harder Pro benchmark splits.

YANAMI introduces a 400K-dialogue multimodal dataset and a 400-dialogue benchmark for multiturn stateful interaction, where models must track, update, and visually realize evolving dialogue states.

EvoTeam is a collaboration-centric evolution framework that maintains reusable agent teams and jointly evolves team organization, team-specific interaction rules, and individual agent skills to improve LLM multi-agent systems across heterogeneous tasks.

VidUEU-Agent is an automated agent pipeline that converts raw videos into 100K multimodal instruction samples and improves understanding, editing, and unified benchmark scores after fine-tuning.

WebHarbor is a suite of 153 dockerized, executable reconstructions of real websites, built through task-driven human-agent collaboration, providing 3,518 tasks for evaluating and RL-training GUI web agents under deterministic control.

ParetoSaddler separates intermediate Pareto-aware promotion from final deployment selection, improving quality and reducing cost on ClawMark and GAIA2.

SpecPilot is an autonomous agentic loop that uses a ceiling-based cost model, parallel coding agents, stock-calibrated token forcing and a structured memory to discover lossless algorithms and systems optimizations for speculative decoding deployments, raising speedup over AR from 4.46× to 10.87× with one request and from 1.94× to 3.93× with batching.

MolRSI is an AI agent that combines experience-guided workflow search with persistent cross-task memory to recursively improve molecular discovery workflows, and it is experimentally validated in antibiotic screening and pMHC binder design.

Caper pre-executes each proposed shell command in a snapshotted sandbox, gives an LLM approver a bounded effects report, commits allowed commands, and rolls back blocked ones with a reason.

Mage-VL-Cowork is a 4B hybrid computer-use agent trained through a five-stage curriculum on 318,837 trajectories generated by evolved GUI and terminal harnesses and reusable workflows, reaching 66.0% on OSWorld-Verified with a 50-step budget.

Recursa adapts a foundation model system to a new domain without human experts or a stronger teacher model by bootstrapping a knowledge repository of instructions, skills, and memory from domain materials, then refining it through parallel multi-agent recursive self-improvement, raising all-pass rate on Harvey LAB's Firm Knowledge benchmark from 30.0% to 55.0% with DeepSeek-V4.1-Flash and from 35.0% to 46.0% with Qwen3.8-Flash.

This report proposes a four-dimension taxonomy for LLM-agent systems that perform computer graphics tasks and builds an agentic literature survey system that applies the taxonomy at scale to 350 papers, revealing design trends and open challenges.

Wave Forcing co-designs causal wavefront generation, three-stage training, and Wave Runtime to stream video at one chunk per concurrent forward round after pipeline fill while retaining all denoising updates.

CUA-Master is a nested double-loop recursive framework that couples open-world seeding and primitive-based task synthesis with iterative model training to improve computer-use agents.

CoworkOS is a cowork-agent harness that separates model-chosen task strategy from runtime execution management and supports long-horizon, human-in-the-loop work across benchmarks, real applications, and automated post-training.

The paper frames video generation through an agent harness that coordinates production, world-model interaction, and development toward recursive self-improvement.

uni-robot-agent is a unified, modular harness that lets open-weight models perform competitive robot use in simulation and the real world, and the paper links effective robot use to four general capabilities.

SWE-Trail uses prior attempts as collector-side history to choose where to compare candidate actions during coding-agent rollouts, then trains students on the resulting trajectories.

NorthCUA is an integrated harness, experience-production flywheel, and staged-learning pipeline that trains Nano, Mini, and Pro computer-use agents and reports strong GUI grounding, desktop, web, and latency results.

PhotoGuider measures four photographic tokens with compact learned scorers to select 360° view directions, rank and refine crops, output camera parameters, and generate measurement-conditioned explanations.

ReTrace is a scalable evaluation framework that formalizes transition-consistent recovery and uses goal-oriented action replay after controlled error injection to systematically study model-based recovery in computer-use agents.

Order-Only On-Policy Distillation trains a student to match the teacher’s ordering of top-K tokens at student-generated prefixes, avoiding teacher probability values and temperature tuning.

IdeaTaste benchmarks LLM research taste by comparing proposals written from a 2025 literature snapshot with papers human communities published in 2026, finding LLM agents match citations and ideas far below human levels and mainly fail to turn directions into concrete approaches.

Reviso is an agentic video captioning framework that learns a Context-Focus-Points observation plan and uses verification-guided SFT plus GRPO rewards to gather prioritized visual evidence with tools.

The paper argues that embedding research across modalities is converging on shared principles and proposes a roadmap from transformation and alignment to joint omnimodal embedding, graded by coverage, comparability, and compositionality.

CrossSWE is a 171-instance benchmark for cross-repository joint development, where tasks require coordinated changes across interdependent repositories and are verified by joint-necessity testing.

The paper finds that frozen video-world-model representations expose modest physical-label information and action sensitivity, but native prediction errors are strongly affected by appearance and static content and can rank impossible events below possible ones.

The paper introduces ReGap-Bench, a paired benchmark that holds scene, question, and answer fixed while varying visual references, and finds that task-matched references improve multimodal reasoning by up to 25% while mismatched references give little benefit or hurt.

FLAMIA is a family of Blackwell attention kernels with MXFP8 and mixed NVFP4/FP8 training paths that use token-block scaling, causal-leakage safeguards, forward-backward numerical consistency, and GPU-side scheduling of packed documents, achieving 1.10–1.40× (B200) and 1.06–1.42× (B300) speedups on combined forward and backward attention calls.

The paper proposes Avatar Intelligence: persistent personal agents that represent a user's goals, values, and constraints, with a user-centric architecture, an L0-L3 delegation roadmap, and the Agentic Oversight Index for evaluating delivered value against human oversight burden.

DecisionRank trains a single-pass decision reranker by combining offline RankNet supervision with item-space on-policy distillation from a generative teacher that provides conditional feedback over candidate items.

FrameForge adapts a frozen video diffusion model to unified conditional image generation by selecting input-adaptive temporal RoPE experts and inserting a latent reasoning chain between condition and target.

R2Orch nests self-play orchestration with a learnability-based Proposer and randomized skill-withdrawal tests, improving a Director toward zero-skill deployment across twelve benchmarks.

The paper argues a qualified yes: co-designing a lightweight System-1 predictor with a costly System-2 executor can reduce execution compute toward 1/ρ_u where interactions are costly, unevenly informative, and predictable before execution.

REA is a routed embodied agent runtime that coordinates a slow System 2 brain, a router and fast System 0 policies so that no running policy waits on the brain.

Apriel Voice 27B Preview is an open speech-language model built on Gemma 4 26B-A4B that uses a two-stage speech-adaptation recipe to unify speech understanding, reasoning, conversation, and tool calls while retaining text and image inputs.

SimpleLoRIF uses scalable per-example gradient embeddings to identify and remove benign-looking fine-tuning data whose gradients align with unsafe behavior, reducing jailbreak attack success while preserving capability.

WeFair trains a compact personalized short-video query rewriter by matching human-feedback fidelity to curriculum SFT, consumption-pair DPO, and strict-target GRPO, reducing intent drift and improving online engagement.

LARC performs reasoning in continuous latent space for unified multimodal models, using curriculum SFT and self-evolving RL to improve image generation and enable test-time scaling.

Wizard is an agentic game world model that decouples explicit game-state control from video generation to enable real-time, zero-shot control of scenes, characters, and motions.

HiFOR couples multi-horizon future prediction (short-term state, semantic milestone, goal) with action-guided retrieval from a structured latent world to condition a flow-matching VLA policy during denoising.

ReLift combines frozen 2D tracks and pretrained geometry into lifted 3D trajectories, then uses a feed-forward trajectory corrector trained on simulated lifts to repair errors and complete occluded motion.

OSWAM learns compact imagination tokens from a video DiT as structured start points for a one-step action DiT denoising, achieving competitive planning and manipulation results with 1 action denoising step and substantially lower latency than multi-step world action models.

CEPO is a framework that splits multi-agent prompts into a shared coordinator prompt and agent-specific executor prompts, and gives the underlying prompt optimizer role-specific meta-instructions for each layer, improving four existing optimizers across 11 benchmarks in four domains.

CaPtain is an open post-training pipeline (SFT on synthetic in-harness trajectories, then multi-turn RL with a privileged simulator critic) that turns a 4B open VLM into a code-as-policy robot agent operating over 22 low-level perception and motion primitives, reaching 73.8% on held-out LIBERO.

The paper functionally probes ten frozen visual foundation models across four spatial capabilities and finds that spatial readout, use of history and motion, predictive transfer, and goal-conditioned action utility dissociate across tasks, layers, and prefix lengths.

WorldBridge is a World Action Model that couples separate video and action experts through structured attention and joint flow matching, and its supervision study shows OOD videos raise RoboTwin 2.0 OOD success from 25.07% to 55.90% with sparse action labels further raising it to 70.70%.

GLOW-FM is a one-step latent flow-matching framework for low-light image enhancement that uses Terminal Velocity Matching in a frozen Tiny-VAE latent space and requires a single function evaluation.

CodeSpecWeave synthesizes coding tasks with interdependent specifications and reliable executable tests, producing 4,670 training tasks and a 100-task benchmark.

The paper splits training-inference mismatch into an approximately fixed numerical discrepancy and a self-induced policy lag, then proposes DLC, which centers the discrepancy, clips the lag ratio, and resynchronizes the sampler when the lag ratio's effective sample size falls below one half.

IDM keeps the full internal DMD correction and adds a feature-moment loss on generator EMA features, improving 2-NFE LongText scores on Z-Image-Turbo and Qwen-Image.

A-RSI uses training-log reward-variation and accuracy signals to decide when to replace a pool of programmatically generated multimodal tasks, and reports gains on an eleven-benchmark suite without external training examples or human answer labels.

The paper recasts a benchmark as an empirical measure over item difficulty, formalizes measure mismatch and coverage gaps as two failure modes with diagnostic metrics and theory, diagnoses 25 benchmarks, and proposes BenchFill to fill large difficulty gaps.

LOWN uses a cheap TEXT-ONLY forward to certify the latest image-conditioned predictions and runs the image-conditioned FULL forward only when the certified prefix is too short, retaining 99.6% of LLaDA-V's score at 35.3% of decoding FLOPs.

Youtu-Parsing-Omni is a 5B omni-modal parser that uses one shared text-initialized encoder and one unified schema to convert documents, images, audio, and video into spatially and temporally grounded structured records, with state-of-the-art document parsing and leading open-weight omni-parsing results.

The paper proves for JEPAs that separately trained modalities with shared positive pairs learn the same representation up to a rotation, and shows empirically that good views align better and more efficiently than bad views or next-token prediction.

L3G-World combines an online-updated latent 3D Gaussian memory with a memory-conditioned video DiT to generate spatially consistent worlds in real time.

SpecSlot decouples output detail from global attention by splitting a fixed token budget into spatial anchors and spectral slots, sending only the lowest-radial slot per anchor through global blocks.

WorkFactory constructs executable professional tasks by planning information support and repairing environments through probe execution, then uses teacher trajectories to train the WorkPilot models.

Co-Design Bench evaluates coding agents on 273 real-user application revisions with hidden acceptance criteria, finding that strong models satisfy many criteria per task but fully complete only a minority of tasks.

VibeGame couples an AI-native 2D game engine with an eight-agent adversarial team and training-free experience distillation to turn natural-language prompts into playable game projects.

HairDeck generates stylized hair cards and spring-bone chains from a single image or point cloud by encoding each card as one token and denoising all tokens together with a set diffusion transformer.

WeVisCut connects a stateful script-to-video agent harness, a four-stage pipeline that recovers executable supervision from finished videos, and a single 9B policy trained through SFT, runtime-verified RFT, Guided-OPD, and RCF-DAPO, evaluated on a 100-task benchmark with five metrics.

IntReal introduces a real-world interleaved dataset and a coarse-to-fine expert-model evaluation suite for multimodal interleaved generation.

CoSTA decouples an SDM's generation and synthesis token spaces by extending the generation vocabulary with BPE-compressed speech units, adapting the model, and deterministically expanding generated units back before the unchanged decoder, yielding about 1.7× speech-token generation speed.

WebAnything automatically reconstructs source-grounded executable web environments from real URLs using an Explorer–Builder and independent verifier, yielding 244 artifacts and 12,984 accepted tasks and improving agent benchmarks after SFT.

Glance equips a single resident 9B generator backbone with a lightweight decision adapter and pointer head to rerank candidate passages from one shared prefill and then generate answers with the same backbone.

Soft-MTP removes the noisy residual from the standard multi-token prediction gradient by recovering the KL self-distillation gradient exactly from existing cross-entropy computations, improving NTP, speculative decoding, and zero-shot accuracy without added cost.

AskStream introduces an ask action and a multi-turn benchmark with Ask-F1 and StreamHarness for evaluating when and what streaming video models should ask under video and user uncertainty.

Falcon is a single-pass causal audio-visual target speaker extractor paired with the RealSSA benchmark, achieving strong extraction quality with much lower parameter, computation, and latency costs than causal AV-TFGridNet.

OmniShowcase generates omnidirectional subject-consistent videos from multi-view reference images by combining a 3D-asset-based data synthesis pipeline with surface-normal turntable conditioning reconstructed from the references.

S2 is a training-free offline metric that measures nearest-neighbor overlap between sensory features and end-effector trajectories and correlates with robot task success across model families.

TAO jointly evaluates tool-return readiness, time-decayed KV-cache credit, and token footprint, then uses dynamic programming to select and place waiting agent sessions on GPU nodes.

RealCLI is a paired CLI benchmark that evaluates autonomous execution and scripted human-agent collaboration on matched tasks, revealing low joint task-collaboration performance among frontier agents.

PersonaCUABench holds 74 tasks fixed while varying the requesting user across 44 personas, and across nine backbones persona requests lower mean task completion from 0.689 to 0.441 while user satisfaction and completion diverge.

TauIndianBankBench is a 1,000-task Indian retail-banking benchmark that scores both required procedure steps and final records for tool-using conversational agents.

AR-BED trains a reasoning LLM with information-gain rewards under an externally specified finite Bayesian experimental-design model, improving adaptive questioning and transferring to a clinical dialogue task.

LongZero lets an LLM improve long-context reasoning from zero external data by first writing structured question-evidence-answer specifications, expanding them into long multi-document contexts, adding self-written retrieval conversations, and training with GRPO.

CausGround distills interaction trajectories into reusable action-centered causal knowledge over a factorized world state, then grounds task-relevant preconditions and effects to guide frozen MLLM planners without parameter updates.

V4.1 Flash-1017 shows strong safety improvements over DeepSeek V4 models in evaluation integrity, loss of control, and harmful manipulation, but retains weaknesses under sustained pressure, partial oversight, and assigned deceptive goals.

Injected metaorders elicit identifiable but too-steep impact laws in five neural LOB generators on five stocks, with no credible post-execution relaxation, while replay and kernel baselines behave as designed.

HuGe trains humanoid generalist VLAs on over 2,000 hours of egocentric human and robot data through a three-stage pipeline, using SONIC motion-token actions and data adaptation to achieve state-of-the-art HumanoidArena results and improved real-world G1 pick-and-drop success.

The paper instruments atomic single-token word-replacement rules and a binned logit model to measure instruction capacity and degradation laws, reporting a GPT-4o mini passport with capacities in rules and uncertainties.

OmniSpectra is a wavelength-native masked-autoencoder foundation model that encodes spectra at continuous physical log-wavelength coordinates instead of pixel indices, letting one 12.2M-parameter model pretrain on eight surveys and match or beat larger instrument-specific baselines without resampling.

PreView is a trained auxiliary preprocessor that supplies center-guided cropped and enlarged views plus candidate-specific visual and textual cues to a frozen GUI grounding model, improving accuracy without retraining that model.

A two-stage paradigm where autoregressive pretraining on semantic tokens acquires visual knowledge and a frozen-backbone pixel-diffusion stage renders images, yielding a 3B text-to-image model trained in about 6,600 H100 GPU-hours that scores 0.86 on GenEval and 87.78 on DPG-Bench.