FuguReport

2026-09-25 Method / Network Security / Automated MUD profile creation

From Source Code to Network Profile: Automated and Traceable MUD Profile Generation for IoT Devices

The paper presents AutoMUD, a source-code-driven system for generating Manufacturer Usage Description (MUD) profiles for IoT devices from firmware and software source trees.

2026-09-25 Method / Self-Training / Sampling strategies in self-training

Strategically Diverse Sampling for Self-Training

This paper investigates self-training data construction for large language models, arguing that repeated samples should vary in problem-solving approach rather than merely in surface form or correctness.

2026-09-25 Method / Meta Learning / Adaptive update via representation adaptation

Online Learning via Learned Latent Bayesian Tracking

The paper studies online learning in non-stationary environments where models must adapt from streaming labeled samples under tight computational constraints.

2026-09-25 Method / 3D Scene Understanding / End-to-end 3D scene graph writing

GraphWrit3R: End-to-End 3D Scene Graph Writing

GraphWrit3R is an end-to-end method for generating 3D scene graphs directly from a point cloud, Gaussian splats, or both, outputting a structured JSON graph of objects, attributes, and relationships.

2026-09-25 Method / World Models / Utility-agnostic interface for slate action modeling

Recommendation World Models for Future-State Control

This paper studies how a trained sequential recommender can be wrapped with a decision layer that reasons about the future consequences of its displayed slates.

2026-09-24 Evaluation / Representation Analysis / Recovery of PoS distinctions

Parts-of-Speech as Emergent Categories in SAE Latent Space

This paper studies how part-of-speech (PoS) information is represented in Sparse AutoEncoder (SAE) latent activations extracted from LLaMA-3-8B.

2026-09-24 Task / Question Answering / Multiple-choice QA evaluation

Likelihood Ranking doesn't Scale Like Prompting in LLMs

This paper compares two evaluation protocols for multiple-choice question answering in large language models: direct prompted answering and likelihood ranking of declarative statements derived from the same question-answer pairs.

2026-09-24 Evaluation / Exploration Evaluation / Measuring AI exploration in alien worlds

ExplorationBench: Measuring AI Systems' Exploration in Verifiable Alien Worlds

ExplorationBench is a benchmark for evaluating whether AI systems can acquire new knowledge through exploration rather than recall.

2026-09-24 Method / Continual Learning / Multimodal geometry preservation

Hyperbolic Multimodal Continual Learning: A Closest-Admissible Solution

This paper studies continual learning for hyperbolic multimodal models, where sequential updates can preserve task scores yet still distort the Lorentz geometry that encodes within-modality similarity, cross-modal alignment, and semantic hierarchy.

2026-09-24 Method / Texture Synthesis / Global coherent texture map synthesis

OmniFabric: Coherent UV Space Texture Synthesis for 3D Garment Reconstruction

OmniFabric is a method for generating simulation-ready garment textures from a single in-the-wild clothing image by synthesizing textures directly in 2D sewing-pattern UV space.

2026-09-23 Method / Mixture-of-Experts / MoE architecture for large language model

Hunyuan-A13B Technical Report

Hunyuan-A13B is an open-source Mixture-of-Experts language model from Tencent featuring 80 billion total parameters while activating only 13 billion parameters per token.

2026-09-23 Method / CAPTCHA Design / Hierarchical motion-based CAPTCHA framework

Invisible in Space, Visible in Time: Motion Vision CAPTCHA against GUI Agents

This paper proposes Motion Vision CAPTCHA (MVCAP), a hierarchical CAPTCHA framework that hides target semantics in motion-defined foreground structures that become recoverable only through temporal segregation from a dynamically evolving background.

2026-09-23 Method / Zero Trust / Safety-aware enforcement framework

Safety-Aware Zero Trust Enforcement for IoT and Cyber-Physical Systems

This paper studies how Zero Trust enforcement should be adapted for IoT and cyber-physical systems, where restricting a suspicious component can reduce cyber risk but also remove sensing or control capabilities needed for safe operation.

2026-09-23 Method / Physical World Modeling / Video-action architecture with flow matching

InternW0: A Foundational Physical World Model for Efficient Real-World Interactions

InternW0 is the first model in the InternW physical world model series, designed to connect future visual prediction with continuous robot control for closed-loop interaction in the physical world.

2026-09-23 Method / Model Scaling / Test-time scaling with planning

Planned Test-Time Scaling with Coordinated Reasoning Paths

This paper proposes Planned Test-Time Scaling (PTTS), a framework for reasoning-time scaling that coordinates multiple inference branches instead of sampling them independently.

2026-09-22 Method / Conformal Prediction / Online conformal state exploration

Risk-Aware Online Conformal State Probing

The paper studies a sequential decision-making setting in which a centralized agent must choose both control actions and whether to probe the true state from a robot or edge device.

2026-09-22 Method / Attribution Methods / Nested subset attribution of components

Matryoshka attribution: Learning to attribute language model outputs to representations and weights

This paper frames attribution in language models as learning an ordering over nested subsets of internal components that minimizes a downstream loss.

2026-09-22 Method / Curriculum Learning / Adaptive training with difficulty buckets

Ladders of Thought: A Self-Evolving Curriculum of Progressively Simplified Reasoning Traces

This paper introduces Ladders-of-Thought (LoT), a training framework for improving reasoning in small- and mid-scale language models by combining progressive question rewrites with an adaptive curriculum.

2026-09-22 Method / Parameter-Efficient Fine-Tuning / Sub-expert level adaptation

From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs

This paper studies parameter-efficient fine-tuning for mixture-of-experts (MoE) language models and argues that expert-level adaptation is still too coarse because activated experts are internally sparse.

2026-09-22 Evaluation / Robotics Benchmarking / Physical condition diversity evaluation

RoboTwin-Phys: Do WAMs and VLAs Understand the Physical World?

RoboTwin-Phys introduces physical-condition diversity as an explicit evaluation axis for robot manipulation, extending the 50-task RoboTwin-2.

2026-09-21 Method / Attention Mechanisms / Expressivity analysis of CKDA

Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

This paper analyzes the expressivity limits of Kimi Delta Attention (KDA) and introduces Complex KDA (CKDA), a variant that extends the gate range to [-1,1] and the delta-rule coefficient to [0,2].

2026-09-21 Method / Video Modeling / 3D-aware memory for video consistency

WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

WorldCrafter is a camera-controllable autoregressive video world model designed to preserve scene content over long horizons and across viewpoint changes.

2026-09-21 Method / Data Annotation / Efficient annotation of alignment data

onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction

onPanda is an interactive annotation tool for LLM alignment data and agent trajectories built around token-level correction.

2026-09-21 Evaluation / Vulnerability Assessment / Evaluation via executable probes

MobileCybench: Evaluating Agent Vulnerability Discovery via Executable Probes

This paper introduces a probe-based framework for evaluating vulnerability reports by replaying an agent's exploit and running executable checks of application security properties.

2026-09-21 Method / Latent Space Learning / Learning geometry-native latent space

GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation

This paper proposes the Geometry-Native Autoencoder (GAE), a compact latent space designed as a shared foundation for both 3D perception and generative visual modeling.

2026-09-20 Method / Motion Estimation / Bézier control points for accurate motion

Accurate Motion Estimation with Bézier Control Point for Efficient Frame Interpolation

This paper presents ABC-Inter, a forward-warping video frame interpolation method designed to address motion ambiguity in training data and the common inference-time assumption of uniform linear motion.

2026-09-20 Evaluation / Behavioral Alignment / Evaluating human-like model behavior

Are Human-Aligned Models Models of Humans? A Turing-Test Gap in Preference Alignment

This paper argues that alignment to human preferences is not the same as alignment to human behavior, and formalizes the discrepancy as a "Turing-test gap.

2026-09-20 Evaluation / Benchmarking / Fine-grained image pair difference

MinCU: A Fine-Grained Benchmark for Grounded Minimal-Change Understanding in Image Pairs

This paper introduces MinCU, a benchmark for grounded minimal-change understanding in near-identical image pairs.

2026-09-20 Evaluation / Memory System Evaluation / Performance on coding-related memory tasks

VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks

VibeMemBench is a benchmark designed to evaluate persistent memory systems in coding agents on real repository tasks with executable validation.

2026-09-20 Task / Human Motion Synthesis / Motion generation from audio and text

MoSAT: Human Motion Generation from Spatial Audio and Textual Description

This paper studies human motion generation jointly conditioned on spatial audio and natural-language descriptions, arguing that the two modalities provide complementary information about what action to perform and where and when it should unfold.

2026-09-19 Evaluation / Benchmarking / Controlled benchmark across diffusion models

Optimizers for Diffusion Models: A Controlled Benchmark

This paper presents a controlled benchmark of seven optimizers for diffusion training across four formulations: masked diffusion on text8, uniform diffusion on QM9, DUO on LM1B, and Gaussian image diffusion on CelebA-64.

2026-09-19 Method / 3D Segmentation / Feed-forward 3D semantic segmentation method

PanoSeg3R: Feed-Forward 3D Semantic Segmentation for Panoramic Images with an Automatic Data Curation Pipeline

PanoSeg3R is a feed-forward framework for 3D semantic segmentation from multi-view panoramic equirectangular images.

2026-09-19 Method / Weak Supervision / Weakly supervised segmentation framework

SomaNet: Weakly Supervised Learning for Instance Soma Segmentation in 3D Electron Microscopy with Partial Annotations

SomaNet is a weakly supervised framework for 3D electron microscopy soma instance segmentation designed for settings with partial instance annotations.

2026-09-18 Method / Visual Navigation / Transformer with pose-based attention

Visual Navigation Transformer with Pose Attention

This paper proposes VNT-PA, a visual navigation transformer that represents an environment as a set of depth keyframes indexed by camera pose rather than as a temporally ordered observation history or an explicit map.

2026-09-18 Evaluation / Model Evaluation / Comparative human-machine cognition evaluation

CogGym: Towards Large-Scale Comparative Evaluation of Human and Machine Cognition

CogGym is a scalable framework for comparing AI models and humans on matched cognitive-science experiments rather than standard answer-key benchmarks.

2026-09-18 Evaluation / Benchmarking / Large-scale pretraining in neural data

BrainWideBench: Benchmarking large-scale pretraining and across-animal transfer in multi-region neural recordings

BrainWideBench is a benchmark for evaluating across-animal transfer in multi-region neural recordings, built on the International Brain Laboratory Brainwide Map dataset.

2026-09-18 Method / Multimodal Learning / Task-aware inference framework

Beyond Exact Match: Task-Aware GRPO for Cross-Domain PCBA Visual Question Answering

This paper studies cross-domain visual question answering for printed circuit board assembly (PCBA), where models must reason over fine-grained visual evidence, component semantics, and manufacturing standards across both standards-derived and real-world production images.

2026-09-18 Method / Privacy Auditing / Calibrated re-identification risk assessment

Conformal Privacy Auditing: Calibrated Re-identification Attacks with Statistical Guarantees

This paper introduces Conformal Privacy Auditing (CPA), a release-time privacy auditing framework for text that quantifies document-level re-identification risk under a declared attacker model.

2026-09-17 Evaluation / Conflict Analysis / Resource conflicts in agent systems

ClashBench: Conflicts Leading Agents to Seize and Harm

This paper identifies and formalizes a safety failure mode in privileged agent systems called destructive resource preemption, where an agent resolves a resource conflict by disrupting an incumbent task instead of reporting the conflict or deferring to the user.

2026-09-17 Method / Self-Distillation / On-policy distillation with privileged information

What Does Privileged Information Add to On-Policy Self-Distillation?

This paper investigates the degree to which privileged information (PI) contributes to on-policy self-distillation (OPSD) beyond the baseline gains of cross-mode distillation itself.

2026-09-17 Method / Attention Mechanisms / Hybrid softmax and linear memory attention

Video DeltaNet: A Video-Native Hybrid Attention for Livestream Video Generation

This paper introduces Video DeltaNet (VDN), a hybrid attention architecture for long-sequence video generation that keeps exact Softmax attention for local temporal interactions and boundary-anchor frames while using bidirectional linear memory for distant context.

2026-09-17 Method / Anomaly Detection / Local sparsity framework for anomaly detection

Local Sparsity Enables Unsupervised LLM Safety Detection

This paper studies deployment-time LLM safety detection without unsafe training labels by framing the task as one-class anomaly detection on model activations.

2026-09-17 Method / Active Learning / Sequential view selection and stopping criteria

Towards Active Cross-View Object Geo-Localization

This paper introduces ActiveGeo, a reformulation of cross-view object geo-localization in which a mobile agent sequentially acquires additional viewpoints and decides when to stop, rather than relying on a single fixed query image.

2026-09-16 Evaluation / Survey / Comprehensive multimodal learning survey

From Models to Systems: A Comprehensive Survey of Efficient Multimodal Learning

This survey examines Efficient Multimodal Learning (EML) as a full-stack problem spanning model design, algorithmic compression, and system deployment.

2026-09-16 Method / Policy Optimization / Critic update strategies in PPO

Rethinking Critic Learning in PPO: Understanding and Mitigating Value Flattening

This paper analyzes a failure mode in PPO critic learning for reinforcement learning with large language models, termed Value Flattening.

2026-09-16 Evaluation / Model Evaluation / Comprehensive multimodal reasoning assessment

Can MiniMax-H3 Reason About the Physical World? An Evaluation of Omni-Modal Generative Model

This paper evaluates whether the omni-modal generative model MiniMax-H3 can reason about the physical world when key information is distributed across multiple input modalities rather than fully specified in text prompts.

2026-09-16 Method / Reinforcement Learning / Real-time policy fine-tuning

Reinforcement Learning for Real-Time Vision-Language-Action Policies

This paper addresses reinforcement learning fine-tuning for large vision-language-action (VLA) policies operating under non-negligible inference latency, which otherwise causes distribution shift from stale observations.

2026-09-16 Evaluation / Benchmarking / Cultural values in language models

GYROval: A Robust Benchmark for Cultural Value Orientation in Large Language Models

This paper introduces GYROval, a benchmark for measuring cultural value orientation in large language models along the two Inglehart–Welzel axes using binary contrastive vignettes.

2026-09-15 Method / Physics-Based Synthesis / Physics-grounded video synthesis

PhysStream: Streaming Physics-Grounded Video Generation with Structured Scene Memory and Fine-Grained Motion Control

PhysStream is an autoregressive image-to-video generation method for physics-grounded, interactive control in multi-object tabletop scenes.

2026-09-15 Method / World Models / Connecting perception and decision making

World Models for Embodied Intelligence: From Plausible to Controllable to Actionable

This paper is a survey of world models for embodied intelligence that argues they should be judged by how they improve behavior rather than by visual realism alone.