FuguReport

2026-10-16 Method / Parameter Efficient Fine-Tuning / Block-wise Toeplitz parameter adaptation

ToPA: Block-wise Toeplitz Adaptation for Expressive and Efficient Fine-Tuning

This paper proposes ToPA, a parameter-efficient fine-tuning (PEFT) framework that reparameterizes weight updates using a product of block-wise Toeplitz matrices combined with an identity residual connection.

2026-10-16 Method / Video Modeling / Persistent dynamics for scene evolution

Momentum: Persistent Dynamics for Video World Models

This paper studies how video world models can maintain consistent scene dynamics even when objects leave the current camera view and later reappear.

2026-10-16 Method / Memory Mechanisms / Online memory gating with delayed feedback

Learning Personalized Memory Gating from Delayed One-Sided Feedback

This paper investigates personalized memory gating for LLM memory systems at the fact level, addressing which extracted facts should be stored under finite memory budgets.

2026-10-16 Method / Generative Models / Order-aware image generation

PONDER: Learning to Generate in the Right Order

PONDER is a policy-learning framework for masked image generation that learns which patches to generate next and which earlier predictions to remove and regenerate later.

2026-10-16 Evaluation / Bias Evaluation / Comprehensive study of positional bias

On the Position Bias of Large Language Models: A Comprehensive Study

This paper presents a comprehensive study of position bias in large language models across multiple-choice question answering and multi-document question answering tasks.

2026-10-15 Method / Reinforcement Learning / Policy optimization with weighted updates

SCOPE: Semantic Bounded Credit Optimization for Prompt Engineering in Text-to-image Generation

This paper studies reinforcement learning for prompt rewriting in text-to-image generation when the image generator is frozen.

2026-10-15 Method / Robot Assistance / Dual-loop harness for assistance

CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance

CarryOn is a dual-loop harness built around a frozen vision-language-action policy to enable personalized and reliable robotic assistance.

2026-10-15 Method / Decoding / Speculative forcing approach

Speculative Forcing for Efficient Autoregressive Video Generation

This paper proposes Speculative Forcing, an inference framework for few-step autoregressive video generators that treats intermediate clean token estimates as drafts and commits tokens early once they appear to have converged.

2026-10-15 Method / Reinforcement Learning / RL integration into diffusion generation

Reinforcement Learning-driven Diffusion Models: Language, Vision, and Beyond

This paper presents a comprehensive survey of reinforcement learning (RL)-driven diffusion models, examining how RL post-training transforms diffusion denoising into reward-guided trajectory optimization.

2026-10-15 Task / 3D Affordance Localization / Localizing functional regions on 3D objects

BridgeAfford: Bridging Query-Target Gap for Generalizable 3D Affordance Grounding

This paper addresses language-guided 3D affordance grounding, which localizes functional interaction regions on 3D objects from natural language instructions, with an emphasis on generalizing to unseen objects.

2026-10-14 Method / Memory Systems / Multimodal agent memory frameworks

Seeing, Maintaining, and Learning: The Evolution of Multimodal Agent Memory

This paper is a survey and analytical framework for multimodal agent memory rather than a new task-specific algorithm.

2026-10-14 Evaluation / Benchmarking / Long-horizon harness performance

Doc2Harness: Benchmarking LLMs on Long-Horizon Harness Engineering of Coding Agents

Doc2Harness is a two-stage benchmark designed to evaluate how well large language models can build coding-agent harnesses rather than only operate within fixed harnesses.

2026-10-14 Application / Multimodal Model / Multimodal code intelligence application

PrismaCoder: Pushing the Boundaries of Open Multimodal Code Intelligence through Decomposed Rubric Synthesis

This paper introduces PrismaForge, a rubric-based pipeline for multimodal code data curation and reinforcement learning, and uses it to build PrismaCoder-2M, a 2 million-sample open dataset for multimodal code intelligence.

2026-10-14 Method / Diffusion Models / Multi-codebook discrete diffusion approach

Intern Lumina U2: A Multi-Codebook Diffusion Large Language Model for Omni-Visual Understanding and Image Generation

Intern Lumina U2 is a unified diffusion language model that represents visual content as fully discrete multi-codebook tokens and handles visual understanding and image generation within one backbone.

2026-10-14 Method / Video Generation / Agentic generation with spatiotemporal control

Any Time, Any Space: A World-State-Centric Framework for Agentic Video Generation

The paper presents Any Time, Any Space (ATAS), a framework for agentic multi-shot video generation built around a persistent world state.

2026-10-13 Method / Agent Training / Scaling long-horizon agent environments

Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards

This paper introduces Tev-Gym, a synthetic training corpus for long-horizon agents built to provide exact verification and dense, progressive rewards.

2026-10-13 Method / Agent Training / Long-horizon coworking agent training framework

WorkRover: Scaling Long-Horizon Cowork Agent Training Via Recursive Environment Evolution

WorkRover is a framework for training long-horizon cowork agents by jointly addressing environment construction, iterative repair, and reinforcement learning credit assignment.

2026-10-13 Evaluation / Benchmarking / Coding agent misbehavior classification

SWE-MISBEHAVIOR: Benchmarking and Mitigating Boundary Violations in Software Coding Agents

This paper introduces SWE-Misbehavior, a benchmark and evaluation framework designed to diagnose boundary-violating behaviors in autonomous software coding agents.

2026-10-13 Method / Model Architecture / Deterministic decision stack with immutable memory

Jeva: System One Models from First Principles

This paper introduces Jeva, a decision-native architecture for low-latency "system one" decision tasks that is designed around four structural requirements: null awareness, permutation invariance, variable-size candidate sets, and cacheable option computation.

2026-10-13 Method / Attention Mechanisms / Attention-based injection in MM-DiT

QK-Edit: Revisiting Attention-based Injection in MM-DiT for Image and Video Editing

This paper presents QK-Edit, a training-free framework for text-guided image and video editing in multimodal diffusion transformers (MM-DiTs).

2026-10-12 Method / Model Conversion / Dialogue-to-Duplex framework

D2: High-Capability Full-Duplex Spoken Language Models from Half-Duplex Backbones

This paper presents Dialogue-to-Duplex (D2), a framework for converting pretrained half-duplex spoken language models into full-duplex models that can listen and speak simultaneously at fine temporal granularity.

2026-10-12 Method / Reinforcement Learning / Training policies for multi-model collaboration

Compositional RL: Evaluating and Training LLMs for Collaborative Reasoning

This paper argues that language models should be evaluated not only by standalone performance but by how much they improve a collaborative multi-model system.

2026-10-12 Method / Model Optimization / Structured parameter updates

Monarch-Muon: Structure-Aware Muon for Fast and Memory-Efficient Pretraining

The paper proposes Monarch–Muon, which combines the Monarch structured matrix parameterization with Muon-style matrix-aware optimization applied directly to the trainable Monarch blocks.

2026-10-12 Method / Reward Modeling / Video reward model without human feedback

IsoVRBench: Video Reward Modeling Without Human Feedback

The paper introduces IsoVRBench, a benchmark for video reward modeling built from controlled interventions that alter a single property of a generated video while holding the rest fixed, enabling preference labels to be derived automatically without per-pair human annotation.

2026-10-12 Method / Knowledge Distillation / Deep-search distillation with entity familiarity

Avoiding What Your Teacher Already Knows: Entity Familiarity for Deep-Search Distillation

This paper studies teacher-relative difficulty in deep-search distillation, arguing that structurally complex questions can still be easy for a teacher if it already knows the key entities.

2026-10-11 Method / Attention Mechanisms / Cleaning and extracting attention maps

Attention Sinks in VLMs Are Borrowed, Not Seen

This paper investigates visual attention sinks in vision-language models (VLMs), where specific image tokens disproportionately absorb attention regardless of query content and distort attention maps.

2026-10-11 Method / Adversarial Robustness / Physical adversarial attack method

Across Viewpoints and Time: Trajectory-Generalizable Physical Attacks on World Models and Video Understanding Models

This paper investigates a practical physical attack setting where a single printable adversarial texture must remain effective across heterogeneous videos and uncontrolled object-camera trajectories.

2026-10-11 Method / Multimodal Video Recovery / Reference-efficient bitstream recovery framework

Reference-Efficient Multimodal Bitstream-Corrupted Video Recovery via Depth-Conditioned Adaptation

This paper studies multimodal bitstream-corrupted video recovery using RGB video together with an auxiliary depth stream that is assumed to be protected and temporally aligned with the corrupted RGB.

2026-10-11 Method / World Models / Symplectic Hamiltonian latent state modeling

LeHami: Symplectic Latent Dynamics for World Models

LeHami is a JEPA-based world model that replaces the standard unconstrained autoregressive predictor with an action-conditioned Hamiltonian flow over a structured latent state split into position-like and momentum-like components.

2026-10-11 Method / Evolutionary Algorithms / Quality diversity test suite search

Small but Mighty: Test Suites as Improvement Signals for Self-Evolving Agents

This paper studies how to generate compact but effective test suites that can serve as improvement signals for self-evolving agents, which must decide whether to keep self-written code rewrites.

2026-10-10 Method / Knowledge Retrieval / Adaptive rule assignment for queries

Co-adaptive Rule Routing and Constraining for Structured Knowledge Retrieval

This paper studies structured knowledge retrieval when the reasoning rule connecting a query to supporting evidence is not labeled in training data.

2026-10-10 Method / Data Selection / Probabilistic relaxation for bilevel selection

Probabilistic Multi-Source Data Selection via Learnable Categorical Priors: Variance Reduction and Guaranteed Derandomization

This paper studies multi-source bilevel data selection, where a model must choose limited subsets from several heterogeneous data sources under per-source budgets.

2026-10-10 Evaluation / Security Benchmarking / Comprehensive secure code evaluation

Last Secure Code Benchmark: Is the Code You Vibe Secure?

This paper introduces the Last Secure Code Benchmark (LSCBench), an evaluation framework designed to assess whether coding agents produce secure as well as functional code in realistic 'vibe coding' workflows.

2026-10-10 Method / Vision-Language Models / Unified multimodal generation

Understanding Unified VLMs: From Attention Probing to Guidance Tuning for Multimodal Generation

This paper studies unified vision-language models that handle both text-to-image and image-to-text generation with a shared Transformer backbone.

2026-10-10 Task / Agent Task Design / Long-horizon multi-skill tasks

SkillHorizon: Constructing Long-Horizon Agent Tasks through Skill Composition

SkillHorizon is a benchmark for evaluating long-horizon agents by constructing executable business workflows through the composition of reusable skills.

2026-10-09 Method / Image Editing / Instruction-guided slider control

SliderEdit++: Improved Slider Trajectories in Instruction-Based Image Editing

This paper studies slider-based control for instruction-guided image editing, focusing on whether slider movements produce smooth, monotonic, and independently controllable edits.

2026-10-09 Task / Representation Learning / Representation quality in DiT

Why DiT Models Underperform as Representation Learners without Long Skip Connections

This paper investigates why Diffusion Transformers (DiTs) produce weaker intermediate representations for downstream perception tasks than U-Net diffusion models, despite their superior generative capabilities.

2026-10-09 Method / Visual Reasoning / Structured intermediate representation for CAPTCHA

CaptchaAgent: Verifiable Visual Reasoning for CAPTCHA Solving

CaptchaAgent is a structured visual reasoning framework for interactive CAPTCHA solving that inserts an explicit visual intermediate representation between page observation and action prediction.

2026-10-09 Method / Optimal Control / Inference-time control for flow models

ConicFlow: constrained Inference-time optimal control for flow matching models

ConicFlow formulates constrained sampling from pretrained flow matching models at inference time as an optimal control problem that trades off terminal reward against the energy required to deviate from nominal generative dynamics, subject to hard terminal constraints.

2026-10-09 Method / Reinforcement Learning / Source-aware credit assignment paradigm

SOAR: Source-Aware Reinforcement Learning with Precise Credit Assignment for Search Agents

This paper addresses evidence fabrication in reinforcement learning for search agents, where models produce reasoning or answers not grounded in retrieved information.

2026-10-08 Method / Video Generation / Goal-directed video world model

WorldGuide: Goal-Directed Video World Model for Procedural Task Execution

The paper formulates procedural video generation as closed-loop task execution in visual world space rather than open-loop synthesis.

2026-10-08 Method / Graph Anomaly Detection / Generalist model without retraining

Is Real-World Training Data Necessary for Generalist Graph Anomaly Detection?

This paper examines whether generalist graph anomaly detection can be trained without real-world anomalous graphs by relying entirely on synthetic data.

2026-10-08 Method / Self-Improvement / Recursive safety improvement framework

ReSI: Recursive Safety Improvement toward Resistant and Resilient AI

ReSI is a recursive safety improvement framework that treats model safety alignment as an iterative loop of red-teaming, automated recipe search, and validated model updates.

2026-10-08 Method / Visual Skill Learning / Framework for visual-native skills

ViSkill: Reinforcing VLM Agents with Evolving Visual-Native Skills

ViSkill is a reinforcement learning framework for vision-language model agents that stores successful interactions as composite visual skill cards rather than reducing them to text.

2026-10-08 Method / Audio-Visual Generation / Simultaneous audio-video waveform generation

SepGen: Multi-Stem Audio-Video Separation and Generation in a Single Model

SepGen extends a pretrained joint audio-video generator so that a single sampling run can produce a video, its mixed soundtrack, and separate waveforms for each captioned source.

2026-10-07 Method / Model Optimization / Token pruning strategies in VLMs

Understanding and Mitigating Token-Pruning-Induced Vulnerabilities in VLMs

This paper investigates how visual token pruning affects the safety of vision-language models under multimodal jailbreak attacks.

2026-10-07 Method / World-Action Models / Causal prediction with historical and future structure

Long-WAM: Scaling the Context of World-Action Models

Long-WAM is a model-and-systems framework for scaling the visual context of causal world-action models under real-time robot control constraints.

2026-10-07 Method / Hand Tracking / Monocular egocentric hand pose estimation

RLHND: Video Foundation Models as Physically Grounded Hand Trackers for Robot Learning

RLHND is a hand tracking method for monocular egocentric video that uses a pre-trained video diffusion backbone as a deterministic feature extractor and predicts both 3D bimanual hand motion and dense tactile signals on the hand surface.

2026-10-07 Method / Memory Mechanisms / Artifact-grounded memory layer

ExperienceIndex: Artifact-Grounded Memory

This paper introduces ExperienceIndex (ExpIdx), a memory layer for AI agents that stores and reuses artifact-grounded experience extracted from prior reasoning traces.

2026-10-07 Method / Graph Neural Networks / Long-range graph learning via operator-valued waves

Oscillatory Neural Dynamics over Sheaves

This paper introduces ONDA, a sheaf neural network framework for long-range graph learning that combines second-order oscillatory dynamics with learned matrix-valued sheaf transport operators.