FuguReport

2026-11-05 Task / Link Prediction / Universal link prediction across graphs

AtlasULP: Domain-aware Universal Link Prediction via Relation Atlas

AtlasULP addresses link prediction on unseen graphs without target-specific fine-tuning.

2026-11-05 Method / Interpretability / Mechanistic interpretability framework

Understanding AI from the Inside: The State of Mechanistic Interpretability for AI Safety

This survey examines whether understanding neural networks' internal representations and computations can improve AI safety, focusing mainly on transformers.

2026-11-05 Method / Visual Interaction / Closed-loop evaluation framework

Interactive Visual Gym: Evaluating Visual Interaction Models with Simulated Users in Interactive Worlds

Interactive Visual Gym evaluates visual interaction assistants in closed-loop simulations where simulated users perform physical or digital tasks and adjust their actions in response to guidance.

2026-11-05 Method / 3D Reconstruction / Streaming high-density reconstruction

MetricTrail: Revisiting the Power of Metric Depth and Local Optimization for Long-Sequence Streaming 3D Reconstruction

MetricTrail is a streaming system for estimating camera trajectories and reconstructing dense geometry from long monocular videos.

2026-11-05 Method / Weight Space / Programming with neural network weights

WeightLoop: Weight-Space Programming with Experimental Feedback

WeightLoop uses a language model to construct a reusable program that turns experimental measurements into bounded, persistent updates to a predictor's parameters.

2026-11-04 Evaluation / Benchmarking / Systematic evaluation of memory-dependent behavior

Robocasa-Mem: Benchmarking Environment and Process Memory in Long-Horizon Mobile Manipulation

RoboCasa-Mem is a benchmark of nine household mobile-manipulation tasks designed to test decisions that require information no longer available in the current observation.

2026-11-04 Evaluation / Model Evaluation / Bayes error as benchmark for overfitting

Practical Estimation of the Bayes Optimal Fairness-Accuracy Tradeoff with Soft Labels

The paper studies the lowest classification error achievable under demographic parity or equal opportunity constraints.

2026-11-04 Application / Video Understanding / Omni video benchmark evaluation

SleuthVLM: Do We Really Need Audio-in-Video for Omni Video Understanding?

SleuthVLM is an agent harness that enables a vision-language model to answer audio-visual questions without directly hearing audio.

2026-11-04 Theory / Sampling Theory / Sampling complexity under smoothness assumptions

Towards the Query Complexity of Sampling under Strong Log-Concavity and Log-Sobolev Inequalities

The paper studies the number of exact potential-value and gradient queries needed to sample smooth distributions on Euclidean space to a specified total variation accuracy.

2026-11-04 Method / Object Detection / Annotation-free detection method

Annotator-Free Object Detection via Foundation Model Collaboration

The paper presents an object detection pipeline that generates training labels without human annotation of the target dataset.

2026-11-03 Method / Reinforcement Learning / Interactive HTML generation training

Catalyzing Interactive HTML Generation Capabilities via High-Fidelity Reinforcement Learning

WebCatalyst is a browser-interaction reinforcement learning framework designed to improve interactive HTML generation from natural-language specifications.

2026-11-03 Theory / Causal Inference / Individualized causal effects modeling

From Type to Token: Petri Nets, Individual Token Philosophy, and Individualized Causal Effects

This theoretical paper proposes Petri nets under the Individual Token Philosophy as a representation of causal mechanisms and counterfactuals for specific individuals.

2026-11-03 Task / Image Retrieval / Retrieval with reference image and text

Factorized MLLM Embeddings for Composed Image Retrieval

Composed image retrieval seeks a target image using a reference image and a text describing the desired change.

2026-11-03 Application / Multimodal Dialogue / Stateful multiturn dialogs

YANAMI: A Large-Scale Multimodal Dataset and Benchmark for Multiturn Stateful Interaction

YANAMI introduces a dataset and benchmark for multimodal dialogues that require models to track and update information across turns.

2026-11-03 Method / Test-Time Adaptation / Open-set test-time adaptation method

Uncertainty-Aware Prototype Learning for Open-Set Test-Time Adaptation

Open-set test-time adaptation must handle unlabeled streams containing both shifted known-class samples and shifted unknown-class samples.

2026-11-02 Method / 3D Reconstruction / Scene mesh recovery from partial point clouds

AnchorComplete: Recovering Complete Scene Meshes from Partial Point Clouds

AnchorComplete reconstructs complete scene meshes from partial point clouds, including geometry outside the observed views.

2026-11-02 Method / Generative Models / Protocol-guided fusion framework

ProtoFusion: Protocol-Guided Prompt Refinement for Generative Infrared and Visible Image Fusion

ProtoFusion combines infrared and visible images using a generative model guided by an explicit, editable fusion protocol.

2026-11-02 Method / Protein Design / Search optimization under structural evaluation

ProteoCraft: Protein Design as Proposal-Verification Search

ProteoCraft treats protein design as a search over discrete sequences under expensive structural evaluation.

2026-11-02 Method / Model Unlearning / Adaptive localization of knowledge negation

Adaptive Localization of Knowledge Negation for Continual LLM Unlearning

The paper studies continual unlearning, where a language model must respond to successive requests to remove different pieces of knowledge without losing general utility.

2026-11-02 Method / Uncertainty Estimation / Specification-aware uncertainty decomposition

Specification-Aware Predictive Uncertainty Decomposition

This paper studies predictive uncertainty when an observed input may incompletely specify an unobserved intent.

2026-11-01 Method / Strategy Optimization / Experience-driven self-evolution framework

SPAR: Experience-Driven Strategy Evolution in Multi-Session B2B Sales

SPAR is an LLM-based framework for improving B2B sales strategies across multiple interactions without updating model parameters.

2026-11-01 Method / Agent Systems / Unified taxonomy for agent systems

Advances in Agentic Computer Graphics

This survey examines graphics systems in which language models choose actions, operate graphics environments, and use execution feedback to guide subsequent steps.

2026-11-01 Method / Robot-Use Agents / Unified harness supporting training

Democratizing Robot-Use Agents: A Unified Harness with Open Models

The paper introduces uni-robot-agent, a modular harness that connects open-weight models to simulated and physical robots through shared tools, environment adapters, and structured feedback.

2026-11-01 Method / Neural Operator / Mesh-to-shell neural operator framework

Mesh2Shell Neural Operator for Inverse Shape Optimization

This paper presents a category-agnostic inverse shape optimization framework that separates an explicit design mesh from the volumetric representation used for physics prediction.

2026-11-01 Method / Scaling Laws / Predictive scaling with hints

Hinted Scaling Laws

The paper addresses the difficulty of forecasting benchmark accuracy when smaller models score near zero and provide little signal for conventional scaling laws.

2026-10-31 Theory / Latent Variable Models / Convergence rate characterization

The Statistical Geometry of Latent Recovery in Hidden Markov Models

The paper studies how accurately latent quantities can be recovered in multivariate Gaussian hidden Markov models when emission centers are close.

2026-10-31 Method / Self-Distillation / Reachability-aligned self-improvement

Steer, Don't Tell: Breaking Reasoning Plateaus with Reachability-Aligned Self-Distillation

The paper investigates how to improve reasoning models on problems that remain unsolved after reinforcement learning, relying only on reference solutions rather than external expert models.

2026-10-31 Method / Dialogue Management / Agentic context management techniques

Agentic Context Management for Multi-Turn Human-AI Conversations

The paper studies context pollution, in which an assistant's earlier interpretations bias later responses even after the user clarifies the task.

2026-10-31 Evaluation / Video Understanding Benchmarking / Ultra-long video QA evaluation

DramaVQA: Probing Detailed Understanding of Ultra-Long TV Dramas

DramaVQA is a benchmark for detailed video question answering across long-form television dramas.

2026-10-31 Method / Memory Learning / Prediction-guided memory update in manipulation

PREM: Prediction-Regulated Memory for Robot Manipulation

PREM addresses partially observable robot manipulation by maintaining a fixed-size recurrent visual memory within a vision-language-action policy.

2026-10-30 Method / Motion Foundation Models / Scalable understanding and generation architecture

Movens-Zero: Scaling Humanoid Motion Foundation Models with Large-Scale Human Videos

Movens-Zero investigates whether large-scale Internet videos can support multimodal humanoid motion generation and closed-loop execution.

2026-10-30 Method / Multimodal Modeling / Unified multimodal understanding and generation

Latent Reasoning in Continuous Space for Unified Multimodal Models

LARC addresses an identified mismatch between discrete language reasoning and continuous image generation in unified multimodal models.

2026-10-30 Task / Scene Text Recognition / Robust text recognition in images

FlexReader: Towards Adaptive Decoding Paths for Efficient Arbitrary-Length Scene Text Recognition

FlexReader is a scene text recognition model designed to vary its decoding stride according to the input.

2026-10-30 Task / Graph Testing / High-dimensional subgraph signal detection

On Statistical-Computational Gaps in Graph Combinatorial Testing Problems

The paper studies strong detection of signals supported on unknown subgraph edges in high-dimensional hypothesis testing.

2026-10-30 Method / Premise Selection / Dependency-aware neural premise ranking

DAPS: Dependency-Aware Premise Selection for LLM Theorem Proving

DAPS is a premise selector for Lean 4 theorem proving that uses dependency structure to identify potentially useful lemmas.

2026-10-29 Evaluation / Semantic Consistency Evaluation / Semantic preservation evaluation in category discovery

Semantic Mismapping in Generalized Category Discovery: Evaluation and Mitigation

Generalized category discovery aims to recognize known categories and discover new ones in the same unlabeled dataset.

2026-10-29 Method / Unified Training / Single algorithm for multiple training stages

Training Foundation Models with a Single Algorithm

The paper formulates pretraining, midtraining, supervised fine-tuning, and suitable reinforcement learning objectives as likelihood fitting over an evolving mixture of static and generated data.

2026-10-29 Method / Vision Transformer / Pixel-space sequence modeling

Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers

SpecSlot is a pixel-space rectified-flow transformer designed to separate output detail from the length of the global attention sequence.

2026-10-29 Task / Knowledge Graph Completion / Predicting missing triples

From Local Recovery to Unified Generation: Discovering Triple Set in Knowledge Graphs via Observation-Anchored Latent Diffusion

Triple set prediction aims to recover a set of missing knowledge-graph facts rather than predict isolated links.

2026-10-29 Method / Autoregressive Models / Generation with arbitrary scan order

Unlocking Any-Order Generation in Pretrained Autoregressive Image Models

This paper presents a method to adapt pretrained raster-scan autoregressive image generators to generate tokens in arbitrary orders.

2026-10-28 Method / Mixture-of-Experts / Efficient hybrid MoE transformer design

Nemotron 3.5 Super and Lightning: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Models for Agentic AI

The paper presents two open hybrid Mamba-Transformer mixture-of-experts models for reasoning and agentic workloads: the multimodal Nemotron 3.

2026-10-28 Method / Stochastic Optimization / Asynchronous non-Euclidean optimization

Rennala NSGD: Asynchronous Stochastic Optimization Beyond Euclidean Geometry

The paper analyzes minibatch stochastic gradient descent for smooth nonconvex objectives in finite-dimensional spaces with arbitrary update norms.

2026-10-28 Method / Test-Time Training / Parallel training of deep nonlinear memory

ParaTTT: Parallelizing Test-Time Training Layers with Deep Memories

ParaTTT parallelizes test-time training layers whose recurrent memory is a neural network updated online by gradient descent.

2026-10-28 Method / Multimodal Learning / Unified multimodal model architecture

Continual Adaptation Across Modalities: Learning and Forgetting in Unified Multimodal Models

This paper investigates continual learning and forgetting dynamics across four unified multimodal models (UMMs) with varying degrees of parameter sharing and different image-generation mechanisms.

2026-10-28 Method / Transformer Models / Operator transformer with heterogeneous experts

HotPOT: Operator Transformer with Heterogeneous Physics Experts for Large-Scale PDE Pre-training

HotPOT is an operator transformer designed for joint pre-training across heterogeneous partial differential equation (PDE) systems.

2026-10-27 Method / Reinforcement Learning / World dynamics action reasoning framework

ThinkWorld: World Imagination Reinforces Chain-of-Dynamics Reasoning

ThinkWorld is a vision-language-action framework that represents action reasoning as a Chain-of-Dynamics (CoD), a sequence of discrete motion tokens intended to capture how the world evolves from the current state toward a future one.

2026-10-27 Method / Autoregressive Modeling / Visual token generation beyond GT

Visual Autoregressive Generation with Next Token Imagination

This paper argues that standard visual autoregressive image generation is overly constrained because training supervises each position only with the ground-truth next token, even though multiple plausible continuations may exist.

2026-10-27 Method / Variational Autoencoder / Compact visual VAE architecture

WithinVAE: Compact Visual Autoencoding with Self-Distilled Semantics

This paper investigates whether semantic structure can be learned directly inside a visual autoencoder without relying on external pretrained recognition encoders.

2026-10-27 Method / Query Reformulation / Multi-query generation and selection

TaskBridge: Technical Report for SIGIR 2026 eCommerce Challenge

TaskBridge is a catalog-anchored query reformulation framework for task-oriented e-commerce search, designed for the SIGIR 2026 eCommerce Challenge.

2026-10-27 Method / Mixture-of-Experts / Channel selection without calibration data

Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data

This paper examines channel selection for mixture-of-experts (MoE) decoding and shows that calibration data are unnecessary for deciding which expert channels to compute.