FuguReport

2026-09-10 Evaluation / Overfitting Analysis / MoE overfit under repeated data

Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data

This paper studies how repeating training data affects sparse Mixture-of-Experts (MoE) language models compared with dense Transformers under fixed total token budgets.

2026-09-10 Evaluation / Benchmarking / Multimodal long-horizon research tasks

Mr.LHDR: A Benchmark for Multimodal Real-World Long-Horizon Deep Research Agents

Mr.

2026-09-10 Method / Video Generation / Real-time interactive video generation techniques

Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation

Vidu S2 is presented as a unified framework for real-time interactive video generation and real-time video editing, comprising Vidu S2-Avatar and Vidu S2-Editing.

2026-09-10 Evaluation / Conversational Agent Benchmarking / Proactive mediation dialogue evaluation

ProMediConv: Benchmarking Proactive Conversational Agents in Legal Dispute Mediation

ProMediConv is a benchmark for legal dispute mediation that models mediation as a proactive, multi-stage, multi-party dialogue process rather than a simple reactive conversation.

2026-09-10 Task / Multimodal Grounding / Linking comments to evidence

ReGround: Grounding Reviewer Comments in Multimodal Evidence

This paper introduces ReGround, a dataset for grounding reviewer comments to specific evidence in original anonymous scientific paper submissions.

2026-09-09 Method / Graph Neural Networks / Theoretical foundations of VNN

Learning with Covariance Matrices: Principal Component Analysis Meets Learning with Graphs

This article presents a graph signal processing perspective on learning with covariance matrices, linking principal component analysis (PCA) to graph-based learning by treating covariance matrices as graph shift operators.

2026-09-09 Task / Multilingual Reasoning / In-language reasoning consistency

Building Multilingual Bridges: Data Mixing as the Pillar of Generalization for In-Language Reasoning

This paper investigates L2 reasoning, defined as a model generating its intermediate reasoning traces in the same language as the user's prompt rather than defaulting to English.

2026-09-09 Method / World Model / Separation of observation and state evolution

Programmable World Model

Programmable World Model is a framework for interactive video generation that separates world-state evolution from visual observation synthesis.

2026-09-09 Method / Visual Token Pruning / Training-free pruning framework

TRACE: Trajectory-robust Admission with Evidence Ordering for Efficient GUI Agents

This paper studies visual token pruning for GUI agents when screenshots are reused through the cache across multi-step trajectories.

2026-09-09 Task / Entity Linking / Linking text and images to knowledge base

Think Before You Link: Rarity, Reasoning, and Retrieval in Multilingual Entity Linking

This paper studies multilingual multimodal entity linking on MERLIN and argues that rarity should not be measured only by popularity signals such as pageviews.

2026-09-08 Method / Diffusion Models / Enhancing training convergence and sampling quality

Revisiting Spectral Representations in Generative Diffusion Models

This paper revisits diffusion modeling through the lens of perturbation kernels and argues that diffusion training is closely connected to self-supervised spectral representation learning.

2026-09-08 Application / Depth Estimation / Monocular depth prediction

Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation

This paper revisits Marigold as a diffusion-transformer-based approach for monocular depth estimation by repurposing the open-source image-editing model Qwen-Image-Edit-2509.

2026-09-08 Method / Profile Learning / Constructing minimal sufficient profiles

Less Is Personal: Learning Minimal Sufficient User Profiles for Personalized Language Models

This paper studies personalized language modeling as a minimal sufficient profile construction problem rather than a fixed-length retrieval problem.

2026-09-08 Method / Policy Optimization / Student policy gradient amplification

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

This paper studies weak-to-strong generalization in on-policy distillation, asking how a stronger student model can benefit from a weaker post-trained teacher without being constrained to imitate the teacher’s final policy.

2026-09-08 Application / Video-Language Modeling / Modeling over space, time, and dynamics

Kairos: A Dataset for Fine-Grained Video-Language Modeling over Space, Time, and Dynamics

Kairos is a long-form video-language dataset designed for fine-grained modeling over space, time, and dynamics using time-resolved annotations.

2026-09-07 Method / Process Mining / Agentic process mining framework

From Event Logs to Governed Action: A BlueSky Agenda for Agentic Process Mining

This BlueSky paper proposes event-to-action process mining, an agenda for transforming heterogeneous operational event data into governed action recommendations for process-aware agents.

2026-09-07 Evaluation / Memory Evaluation / Recall capacity benchmarking

On the Recall Scaling Laws in Mamba: A Theoretical and Mechanistic Study via Hashing

This paper examines associative recall in Mamba through mechanistic interpretability to identify the internal algorithm used on multi-query associative recall (MQAR).

2026-09-07 Method / Reinforcement Learning / Reward design for long-horizon tasks

Long-Horizon Language Model Reinforcement Learning via Progressive Point Matching

This paper studies reinforcement learning for language models on long-horizon reasoning tasks, where standard sparse outcome rewards provide weak training signal.

2026-09-07 Method / Invariant Learning / Learning with exact group invariances

Efficient Learning and Symmetry Discovery under Exact Invariances

This paper studies supervised regression under exact group invariances and addresses whether invariant estimators can be computed efficiently when the symmetry group is finite, infinite, or unknown.

2026-09-07 Method / Reinforcement Learning / Training coding agents with RL

FrogNano: Training a 4B Coding Agent via Online Task Synthesis

FrogNano is a 4B-parameter coding agent designed for repository-level software engineering tasks, trained exclusively through reinforcement learning on approximately 1,500 synthetic SWE environments.

2026-09-06 Method / Recursive Improvement / Self-improving system design

MetaRSI / RSI2: A Meta-Recursive Self-Improving System for Recursive Self-Improving Systems Themselves

The paper presents MetaRSI-v1, a meta-recursive self-improvement framework that treats a deployed system as three writable surfaces: data, execution harness, and model parameters.

2026-09-06 Method / Gradient Control / Mitigating interference in updates

One Step, One Lead: Mitigating Higher-Order Interference in Multi-Domain Reinforcement Learning via Cross-Step Control

This paper studies interference in multi-domain reinforcement learning for large language models from a cross-step perspective rather than a single-step gradient view.

2026-09-06 Method / MoE Routing / Routing effective rank definition

From Concentration to Differentiation and Back: Routing Effective Rank in MoE Reasoning Cohorts

This paper proposes routing effective rank, an entropy-based measure of the spectral dimensionality of a cross-rollout similarity graph built from MoE expert-routing traces during test-time reasoning.

2026-09-06 Evaluation / Multimodal Reasoning / Benchmark for cultural interpretation

NormViz: A Benchmark and Framework for Grounding Multimodal Reasoning in Global Cultures

NormViz introduces a human-validated benchmark for visual norm understanding: 3,272 contrastive image pairs spanning 16 countries, where each pair differs only in a culturally relevant behavior that changes the image's interpretation.

2026-09-06 Theory / Representation Theory / Theoretical aspects of feature superposition

Feature Superposition in Neural Networks: From Theory to Practice

This paper is a survey of feature superposition in neural networks, defined as representing more features than dimensions.

2026-09-05 Method / Multimodal Fusion / Text-guided LLM adapter

MVFA: A Multi-View Text-Guided Multimodal Fusion LLM Adapter for Sentiment Analysis and Emotion Recognition

This paper proposes MVFA, a parameter-efficient adapter for multimodal sentiment analysis and emotion recognition in conversations with frozen large language models.

2026-09-05 Method / Online Learning / Adaptive prompt routing with limited feedback

Online Learning with LLM Experts from Limited Feedback

This paper studies online prompt routing among multiple LLM experts when response-quality feedback is available only for a limited number of rounds.

2026-09-05 Method / Skill Routing / Diversity-aware routing framework

Beyond Top-$k$ Skill Retrieval: Diversity-Aware Skill Routing for LLM Agents

This paper studies skill routing for LLM agents when a large registry contains many overlapping or redundant skills and complex tasks require complementary skill sets.

2026-09-05 Method / 3D Reconstruction / Dense-view feed-forward splatting

AVSplat: Dense-View Feed-Forward 3D Gaussian Splatting with Assist-View Preconditioning

AVSplat addresses a failure mode of pose-free feed-forward 3D Gaussian Splatting in which adding many input views can reduce reconstruction quality.

2026-09-05 Evaluation / Visual Benchmarking / Fine-grained image difference evaluation

VDiff-Bench: A Challenging Benchmark for Fine-Grained Image Difference Identification

VDiff-Bench is a multiple-choice benchmark for fine-grained image difference identification in multimodal large language models.

2026-09-04 Method / Domain Generalization / Risk-aware adversarial tuning and generation

PAPT++: Risk-Aware Adversarial Tuning and Generation for Single Domain Generalization

PAPT++ addresses single domain generalization by using a pretrained text-to-image diffusion model to generate training samples that are not only diverse but also difficult for the current classifier.

2026-09-04 Evaluation / Reasoning Assessment / Mathematical reasoning knowledge evaluation

Do LLMs Exhibit Coherent Knowledge Structures in Mathematical Reasoning? A Perspective from Knowledge Space Theory

This paper proposes a Knowledge Space Theory (KST)-grounded framework to evaluate whether large language models exhibit coherent knowledge structures in mathematical reasoning rather than only high answer accuracy.

2026-09-04 Method / Model Explanation / Causal dependency graph construction

DCFA: Dual-view Causal-inspired Attribution for Failure Reasoning in LLM-based Multi-agent Systems

This paper addresses failure attribution in large language model (LLM)-based multi-agent systems, aiming to identify the decisive error whose correction reverses or maximally reduces system failure.

2026-09-04 Method / Agent Environments / Scalable environment pipeline

CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents

CUA-Universe is a pipeline for turning real desktop software into hybrid environments where agents can act through both graphical interfaces and command-line tools over shared application state.

2026-09-04 Method / World Model / Physics-informed world modeling

TourPhysics: Bringing Physics to World Models for Exploration and Manipulation from a Single Image

TourPhysics is an online world-model framework that initializes a controllable scene hypothesis from a single image and a declarative physical configuration.

2026-09-03 Method / Test-Time Adaptation / Adaptation via human-AI interaction

Efficient Test-Time Adaptation through Human-AI Interaction

This paper introduces TAHI (Test-time Adaptation through Human-agent Interaction), a framework enabling interactive AI agents to adapt to individual human expertise on open-ended tasks.

2026-09-03 Method / Attention / Random token preservation in attention heads

Random Attention: Rethinking KV Cache Eviction for Efficient Reasoning

This paper studies KV-cache eviction for long-chain reasoning in large language models and questions the standard assumption that cached tokens must be ranked by an importance score.

2026-09-03 Method / Diffusion Models / Training diffusion transformer from scratch

LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes

LLaDA-Image is a unified image generation framework centered on a 6B Diffusion Transformer trained from scratch and conditioned by a frozen dLLM-based vision-language module.

2026-09-03 Method / Embedding / Distillation for compositional reasoning

CORE: Improving Compositional Reasoning in MLLM Embedding via Reranker Distillation

This paper studies a gap between cross-attentive rerankers and dual-encoder embeddings in multimodal retrieval: rerankers can often make fine-grained compositional judgments that are not preserved in embedding space.

2026-09-03 Method / Pose Estimation / Multi-relative pose query learning

Scal3R: Learning Efficient Multi-Relative Pose Query for Scalable Online 3D Reconstruction

Scal3R addresses the failure of online 3D reconstruction models on long video sequences by replacing absolute pose regression to a fixed first frame with multi-reference relative pose queries.

2026-09-02 Method / Video Modeling / Interactive video world model construction

SolarWM: Open Data and Scalable Training for Long-Horizon Video World Models

SolarWM is introduced as an open foundation for interactive video world models that combines an open data engine with a unified training stack for heterogeneous video backbones.

2026-09-02 Method / Model Fine-Tuning / Post-training with SFT and RL

Post-Training Language Models for Gold-Medal Performance in Coding Competitions

This paper presents an end-to-end post-training and inference pipeline for competitive programming that integrates large-scale problem curation, synthetic reasoning traces, supervised fine-tuning (SFT), reinforcement learning (RL), and an iterative test-time compute procedure called GenCorrect.

2026-09-02 Method / Attention / Controlling attention in decoding

Language Models Can Control Their Own Attention

This paper introduces Declarative Attention (DA), a prompting and inference protocol in which a language model explicitly declares where it needs to attend during decoding.

2026-09-02 Method / Reinforcement Learning / Tail-likelihood based learning

Tail-Likelihood Reinforcement Learning

This paper proposes Tail-Likelihood Reinforcement Learning (TailRL), an RL objective for continuous rewards that maximizes the expected log-probability of exceeding a uniformly sampled reward threshold.

2026-09-02 Method / In-Context Learning / Task-specific reasoning operation learning

SALA: Semantic-Aware Logical Alignment for Complex Reasoning in In-Context Learning

This paper studies demonstration selection for complex reasoning in in-context learning, arguing that surface-level similarity and rigid predefined reasoning-step matching do not adequately capture problem-solving logic.

2026-09-01 Method / Model Training / Training personalized simulators

StudentSim: Training LLM-based Student Simulators

StudentSim is a two-stage training framework for building individualized student simulators from sparse per-student data.

2026-09-01 Method / Reinforcement Learning / Off-policy and on-policy reward supervision

Scaling Near-Optimal SFT-RL Annotation Budget Allocation from Small to Large LLMs

This paper studies how to split a fixed post-training annotation budget between supervised fine-tuning (SFT) and a subsequent RL stage for large language models.

2026-09-01 Method / Multimodal Models / Unified understanding-generation framework

Uncovering Understanding-Generation Synergy in Native Unified Multimodal Models: From Representation, Task to System

This paper studies whether unified multimodal models actually gain from jointly learning visual understanding and image generation, rather than merely exposing both functions through one interface.

2026-09-01 Method / Knowledge Distillation / Mid-training distillation with confident tokens

Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall

This paper studies how logit-based knowledge distillation behaves during mid-training, the continued self-supervised stage between pre-training and post-training.

2026-09-01 Method / Sparse Models / Input-adaptive sparse prefilling technique

CRISP: Cliff-awaRe Input-adaptive Sparse Prefilling with Structural-Mass-Motivated Routing

CRISP is a training-free method for long-context sparse prefilling that targets two limitations identified in prior dynamic sparse attention systems.