Tagged articles
NeurIPS 2026
46 articles · Page 1 of 1
A NeurIPS 2026 study reveals that vision-language models (VLMs) significantly differ from humans in scene understanding: VLMs over-rely on object size, centeredness, and low-level saliency, while humans naturally focus on people. The researchers introduce Counterfactual Semantic Saliency (CSS), a black-box evaluation method, and a dataset of 307 scenes with 1,306 counterfactual variants, showing all 19 tested VLMs fall short of human-level alignment, with size sensitivity being the primary driver of this gap.
Black-Box EvaluationCausal InterventionCounterfactual Semantic Saliency
0 likes · 13 min read

Researchers from Tsinghua and Peking University discover that massive activations in transformer models act as adaptive gradient regulators, with attention sinks creating gradient sinks that RMSNorm suppresses, revealing a backward-propagation mechanism linking attention concentration to activation scaling.
Attention SinkBackpropagationGradient Sink
0 likes · 11 min read

A NeurIPS 2026 paper from Lund University and Google DeepMind introduces a forward-only adaptation method for LLMs that uses angular concentration of SVD components during inference to achieve task-specific adaptation without backpropagation, reducing peak memory and increasing speed compared to gradient-based and exhaustive search baselines.
LLM adaptationNeurIPS 2026SVD
0 likes · 9 min read

A NeurIPS 2026 paper introduces a forward-only adaptation method for LLMs that uses angle concentration of SVD components to prune ranks during inference without backpropagation, achieving lower peak memory and higher speedup than gradient-based and exhaustive search methods across diverse models and tasks.
LLM adaptationNeurIPS 2026SVD
0 likes · 9 min read

Hong Kong University researchers identify implicit advantage symmetry as the root cause of GRPO's poor exploration and difficulty adaptation, proposing A-GRAE to dynamically break symmetry and achieve simultaneous gains in Pass@1 accuracy and Pass@k diversity across math and multimodal benchmarks without extra compute.
A-GRAEGRPOLLM post-training
0 likes · 13 min read

Aalborg University researchers show diffusion language models leak less training data than autoregressive models at equal scale, but increasing sampling steps during generation significantly raises exact memorization extraction risk, as demonstrated on Enron email completion tasks with LLaDA-8B and matched 1.1B models.
Aalborg UniversityLLaDA-8BNeurIPS 2026
0 likes · 10 min read

Aalborg University researchers show diffusion language models leak less training data than autoregressive models at equal scale, but finer sampling steps sharply raise exact memorization risk — LLaDA-8B's extraction success jumps 20× when moving from one-step to token-by-token generation.
Aalborg UniversityLLaDA-8BNeurIPS 2026
0 likes · 11 min read

Tsinghua University's SkillMaster framework teaches LLM agents to create, update, and retain reusable skills through reinforcement learning, using counterfactual utility rewards and DualAdv-GRPO to jointly optimize task execution and skill management, achieving 98.7% on ALFWorld and 82.0% on WebShop with Qwen2.5-7B.
DualAdv-GRPOLLM agentsNeurIPS 2026
0 likes · 11 min read

Aalborg University researchers present a generalized extraction framework showing diffusion language models leak less training data than autoregressive models at same scale, but finer sampling steps significantly increase exact memorization recovery, with LLaDA-8B showing a 20x increase in email extraction when switching from one-step to token-by-token generation.
LLaDANeurIPS 2026autoregressive models
0 likes · 11 min read

HKUST MMSense Lab introduces a visual-native agent framework that assigns unique IDs to images for tool chaining in multimodal deep search, and an on-policy data evolution method that iteratively refines question-generation rules using agent rollouts, boosting 8B and 30B model accuracy across eight benchmarks.
HKUST MMSense LabImage ReuseMultimodal Deep Search
0 likes · 20 min read

Researchers from Northwestern Polytechnical University and HKUST propose Dynamic Gated Pathways (DGP), a method that lets each LLM layer dynamically attend to all ViT layers, improving multimodal benchmarks by learning per-layer, per-sample routing weights instead of using a fixed visual projection.
DGPDynamic Gated PathwaysLLM
0 likes · 18 min read

Touch-R1 introduces the first tactile-anchored physical reasoning model using reinforcement learning to enable cross-sensor tactile inference, correcting visual misjudgments by prioritizing tactile evidence, and achieves state-of-the-art results on the new TouchReason-Bench benchmark across nine metrics including physical property identification, ordinal reasoning, and cross-sensor consistency.
Cross-Sensor ConsistencyEmbodied AINeurIPS 2026
0 likes · 10 min read

The paper introduces LCVN, a language-conditioned visual navigation task where models predict complete action sequences from a single starting image and a language instruction, comparing a world-model-plus-policy pipeline against a unified autoregressive model on a new dataset with three instruction styles.
LCVNNeurIPS 2026autoregressive models
0 likes · 11 min read

Firefly introduces a back-chaining pipeline that explores real MCP servers to generate verifiable tool-call tasks, builds a retrieval-augmented simulator for reproducible offline RL, and shows a 4B model reaching Claude-level performance on tool-use benchmarks.
Back-ChainingData SynthesisMCP Servers
0 likes · 27 min read

The Shenzhen Institute of Big Data and Huawei present ASP, an adaptive scheduling pipeline for multi-instance asynchronous RL training on Ascend 910B3 NPUs, eliminating padding, skewness, and sorting bubbles to achieve 222% higher generation throughput and 76% lower iteration time, validated at NeurIPS 2026 and deployed in telecom cutting training from three months to two.
Ascend NPULLMMindSpeed-RL
0 likes · 20 min read

HKU researchers propose RaPO, a reinforcement fine-tuning method that adds a retention reward based on trajectory-level drift to prevent catastrophic forgetting in multimodal large language models during visual continual learning, achieving state-of-the-art results on five benchmarks with minimal forgetting.
Catastrophic ForgettingContinual LearningGRPO
0 likes · 16 min read

ReSFlow maps categorical probabilities onto a sphere and rectifies generation paths along Fisher–Rao geodesics, achieving one-step generation of discrete data like images and DNA sequences with quality matching multi-step methods and a 166× speedup.
DNA Sequence DesignDiscrete Data GenerationFisher-Rao Geometry
0 likes · 15 min read

Hong Kong University and Yisheng Technology introduce DexHoldem, a real-robot benchmark using Texas Hold'em to test both dexterous manipulation (1,470 demos, 14 skills, UR10e+Shadow hand) and embodied perception/reasoning (36 questions, 14 model combos), with metrics TCR and SPSR revealing gaps between action success and scene preservation.
DexHoldemEmbodied AINeurIPS 2026
1 likes · 10 min read

MemSkill introduces memory skills as learnable, reusable modules that enable agents to continuously improve memory construction via task feedback and failure analysis, achieving state-of-the-art results on long-term dialogue and embodied interaction benchmarks with zero-shot transfer across models and tasks.
MemSkillNeurIPS 2026Recursive Self-Improvement
0 likes · 11 min read

Researchers from Tianjin University and Alibaba propose MIPU, a reinforcement learning method that addresses training-inference mismatch in LLMs by directly optimizing inference-time performance, achieving higher accuracy and more stable training on Qwen3 models across math benchmarks.
LLM trainingMIPI PrincipleMIPU
0 likes · 11 min read

Tianjin University and Hefei University of Technology researchers propose FAME, a multimodal exploration framework using Conditional Flow Matching to generate multiple exploration anchors, combined with A* path planning, clustering, and hierarchical selection, achieving state-of-the-art coverage on AiMDoom and Matterport3D benchmarks with strong cross-dataset generalization.
Active 3D MappingConditional Flow MatchingCross-Dataset Generalization
0 likes · 12 min read

NV-CoT replaces discrete tokenized bounding boxes with continuous numerical coordinate prediction via four lightweight linear heads and a Laplace policy for reinforcement learning, raising region localization IoU from 47.3 to 59.5 and improving high-resolution VQA accuracy on V* Bench and HR-Bench under both supervised fine-tuning and GRPO.
Continuous CoordinatesMultimodal Large ModelsNV-CoT
0 likes · 11 min read

EchoChat introduces a unified framework for empathetic speech dialogue that decomposes empathy into acoustic perception, Theory of Mind reasoning, and response generation, using Acoustic-Anchored Attention and step-decomposed credit assignment to improve understanding and response alignment, evaluated on new datasets EchoDialogue-400K and EchoEval.
Acoustic-Anchored AttentionEchoChatEchoDialogue-400K
0 likes · 10 min read

Researchers from Shanghai AI Lab and Fudan University introduce ReasoningShield, the first benchmark and moderation models for chain-of-thought safety, showing that final answer safety doesn't guarantee reasoning trace safety, and achieving SOTA on reasoning trace moderation with 91.8% F1.
AI safetyNeurIPS 2026ReasoningShield
0 likes · 15 min read

Researchers from Renmin University, Rutgers, NUS, and MBZUAI introduce TabClustPFN, a prior-data fitted network that performs zero-shot clustering on tabular data via a single forward pass, automatically estimating the number of clusters and providing a probability distribution over candidate cluster counts, achieving competitive performance on 44 real-world datasets without per-dataset training.
Cluster Count EstimationNeurIPS 2026PFN
0 likes · 10 min read

ReasoningShield introduces the first chain-of-thought safety moderation benchmark with 9.2K samples across 10 risk categories, training 1B/3B models that achieve 91.8% F1 on reasoning traces—outperforming GPT-4o (87.8%) and LlamaGuard-4-12B (56.2%)—while localizing risk evidence to specific reasoning steps.
AI alignmentLLM safetyNeurIPS 2026
0 likes · 13 min read

Researchers from HKU and Tsinghua propose VAST, an offline RL framework that adaptively chooses time horizons for value updates by stitching variable-length trajectory segments, improving long-horizon decision-making across 50 OGBench tasks.
Adaptive Horizon SelectionLong-Horizon RLNeurIPS 2026
0 likes · 10 min read

This NeurIPS 2026 paper introduces Span-Level Uncertainty Estimation (SLUE) and SpanUQ, a lightweight probe that detects semantic spans and quantifies their uncertainty in a single forward pass, achieving 10-20x speedup over sampling-based methods while providing fine-grained, actionable uncertainty signals for downstream decision-making.
LLM UncertaintyNeurIPS 2026Probe Model
0 likes · 31 min read

ReGDiff introduces a two-step method — RAS for latent space regulation and SRR for short-range repulsion during sampling — to generate geometrically valid, novel, and diverse metamaterial voxel structures, outperforming baselines on MetaTruss and MetaShell datasets with significant improvements in geometric rationality, novelty, and diversity.
AI for ScienceGenerative AIMetaTruss dataset
0 likes · 14 min read

RVCBench introduces a comprehensive benchmark evaluating 18 voice cloning models across 18 robustness tests covering input shifts, generation challenges, output distortions, and adversarial audio perturbations, revealing that clean-condition performance does not predict real-world reliability and that model architecture matters more than size.
NeurIPS 2026RVCBenchadversarial audio
0 likes · 22 min read

WorldMemArena, a NeurIPS 2026 benchmark from UCSB, Stanford, ETH Zurich, and others, evaluates multimodal agent memory across write, maintain, retrieve, and use stages in dynamic action-world loops, showing high recall rates (86%) don't guarantee correct answers (50%).
Multimodal Agent MemoryNeurIPS 2026WorldMemArena
0 likes · 9 min read

University of Pittsburgh researchers show that vision-language models internally represent 3D scene topology in middle-layer latent spaces, revealed by separating appearance attributes and visualizing spatial structure; Dirichlet energy quantifies neighborhood preservation, and fine-tuning with this loss improves spatial reasoning on benchmarks like VSI-Bench.
3D Scene UnderstandingDirichlet EnergyInterpretability
0 likes · 12 min read

Xiaohongshu and Fudan University propose DR-Rubric, a framework where models perform deep research to gather evidence and synthesize atomic constraints as rubrics for reinforcement learning rewards, achieving competitive results on six benchmarks with only 1,000 training samples while revealing non-monotonic bootstrap dynamics and reward polarization as a stopping signal.
DR-RubricDeep ResearchGRPO
0 likes · 21 min read

STILL introduces a content-aware intra-layer hybrid attention mechanism that routes tokens to Softmax or linear attention based on self-saliency scores computed locally, preserving pre-trained norm structure via NP-Map, achieving 80% memory reduction and 83% prefill speedup with only 0.04B training tokens.
Efficient TransformersIntra-Layer Hybrid AttentionLLM Linearization
0 likes · 20 min read

Fudan and SJTU researchers open-source A3, a reinforcement learning framework that uses AST-based structured credit assignment to train a 14B CLI agent achieving 24.6% exact match on ShellOps, doubling baselines, and outperforming a 235B model on out-of-distribution tasks with minimal computational overhead.
CLI agentsCredit AssignmentNeurIPS 2026
0 likes · 18 min read

FlowMAS, a NeurIPS 2026 paper, uses Generative Flow Networks with curiosity-driven and information-theoretic rewards to automatically generate diverse, efficient multi-agent workflow topologies, outperforming baselines across six benchmarks with far lower computational cost.
Automated Architecture DesignCuriosity-Driven ExplorationGFlowNets
0 likes · 11 min read

An analysis of NeurIPS 2026 acceptance data using the open-source PilotDeck 2.0 agent reveals Tsinghua leads in institutional participation while overseas institutions dominate high-productivity authors, alongside trends showing 3.4x submission growth, Benchmark papers dominating, and Agent research surging 2.42x.
AgentOSNeurIPS 2026PilotDeck 2.0
0 likes · 9 min read

RA-CFGCache introduces a risk-aligned caching framework for classifier-free guided diffusion models that jointly models cross-branch error alignment and timestep-dependent error propagation, achieving superior efficiency-fidelity trade-offs on FLUX.1-dev, Wan2.1, and CogVideoX while remaining compatible with existing cache estimators.
Classifier-Free GuidanceGenerative AIModel Acceleration
0 likes · 19 min read

This article analyzes TokenRouter, a NeurIPS 2026 system that enables efficient token-level LLM routing through a request-centric programming interface, asynchronous model handoffs with state preservation, and throughput-optimized delayed batching, demonstrating 2-64x throughput gains over baselines across multiple routing algorithms and workloads.
LLM servingMarkov chain modelNeurIPS 2026
0 likes · 30 min read

Apple and Duke University propose LensVLM, a vision-language model that compresses long documents into images for efficient browsing, then lets the model selectively magnify relevant pages by retrieving original text, achieving 68.9% accuracy at 4.3x effective compression on seven QA benchmarks, nearing full-text performance while reducing latency and GPU costs.
LensVLMNeurIPS 2026effective compression rate
0 likes · 11 min read


Amap TechOct 12, 2026 · Artificial Intelligence
Amap had 16 papers accepted at NeurIPS 2026, covering multimodal large models, spatial intelligence, visual reasoning, and reinforcement learning, with highlights including a decomposed verifier for generative world modeling, gradient projection for continual multimodal learning, map-free transit route generation, and benchmarks for coding agents and travel planning.
3D reconstructionEmbodied AIMultimodal Large Models
0 likes · 30 min read

TerraVis introduces a systematic framework for evaluating world consistency in AI-generated images, covering 18 issue types across object, interaction, and scene levels, validated on 13,500 human annotations across five models.
AI-generated imagesNeurIPS 2026benchmark
0 likes · 15 min read

Sichuan University researchers propose PAL, a two-stage pairwise artifacts learning framework, and the EditGroup-45K dataset to explicitly decouple manipulation artifacts from semantic features, achieving a 10.3% cross-domain F1 improvement and the first NeurIPS Oral in image manipulation localization.
Artifact DisentanglementComputer VisionCross-Domain Generalization
0 likes · 11 min read

NeurIPS 2026 paper SAGE introduces structural guidance via Symbolic Closure Analysis to fix long-horizon reasoning biases, using algebraic sparsification and hyperbolic distance feedback during training, outperforming baselines on 12 benchmarks and achieving nearly 8x Lean verification pass rate on Andrews-Curtis tasks.
Algebraic SparsificationAndrews-CurtisHyperbolic Space
0 likes · 16 min read

VisInteract introduces a dynamic interactive Text-to-Vis paradigm with the VisInteract-Bench benchmark and Vis-MCTS method, using Monte Carlo Tree Search with progressive widening, cross-path memory, and dimension-aware reward decomposition to recover user intent from imperfect queries, achieving 51–57% Merge Task Score on 1,098 samples across 11 databases.
Imperfect QueriesMonte Carlo Tree SearchNeurIPS 2026
0 likes · 18 min read

Researchers from Virginia Tech, UW-Madison, and Dartmouth introduce SAGE, a post-training method that uses symbolic closure analysis to diagnose exploration and accumulation biases in long-horizon reasoning, applying algebraic sparsification and hyperbolic structure guidance to improve sampling and provide early feedback, achieving near 8x Lean verification pass rate on Andrews-Curtis tasks and outperforming baselines across 12 benchmarks.
Andrews-CurtisLarge Language ModelsLean verification
0 likes · 14 min read
