NeurIPS 2026 spotlights

Best rated first.

92%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightTsinghua Xiaomi Corporation Peking Institute of Automation, Chinese The Hong Kong University of Scie Vision-language-action models

Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation

Embodied chain-of-thought improves generalization when grounded in action guidance, and ERVLA uses reasoning-dropout supervision to avoid unstable autoregressive reasoning at inference, achieving state-of-the-art robot manipulation results.

Nan Sun, Yuan Zhang, Yongkun Yang, Wentao Zhao and 9 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

19/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 19 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 4/5

92%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightOsnabrueck Osnabrück Universität Osnabrück U Osnabrück Bergische Universität Wuppertal Object detection

Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation Uncertainty

Re-annotation reveals object detection benchmarks miss up to 60% of objects due to incomplete labels, and current detectors remain misaligned with human perception.

Sarina Penquitt, Jonathan Klees, Antonia van Betteray, Parssa Jashnieh and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

19/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 19 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 4/5

91%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightMicrosoft U Wisconsin-Madison RL for LLMs

ECHO: Terminal Agents Learn World Models for Free

ECHO trains terminal agents to predict environment responses for dense supervision, doubling GRPO pass@1 on TerminalBench-2.0.

Vaishnavi Shrivastava, Ahmed Awadallah, Dimitris Papailiopoulos

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 7 on Hugging Face

– ReadersNo votes yet. 1 from authors or colleagues not counted

18/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 18 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 3/5

91%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Toronto / Vector Institute Cornell BlankBio Memorial Sloan Kettering Cancer U Toronto Genomics & single-cell

mRNABench: A curated benchmark for mature mRNA property and function prediction

mRNABench benchmarks mature mRNA property predictions across 59 tasks and 135K experiments, revealing synergies between self-supervised objectives that yield a compact state-of-the-art Mamba model using 700x fewer parameters.

Ruian (Ian) Shi, Taykhoom Dalal, Philip Fradkin, Divya Koyyalagunta and 9 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

18/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 18 of 20 reviewers recommend it

lenient 5/5

medium 9/10

strict 4/5

91%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightHunan U Hong Kong Mechanistic interpretability

CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models

CRAFT localizes arbitration and brake failure in medical vision-language models to disjoint attention head sets, enabling targeted interventions that reduce misleading text influence and restore abstention without retraining.

Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen and 2 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

18/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 18 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 3/5

91%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Neuchatel U Neuchâtel Delft University of Technology IBM Research U Turin AI oversight & deception

Safety Reconstructed: Generative Modeling via Masked Diffusion Builds Strong Safety Guardrails

LLaDA-Guard uses masked diffusion to score responses under each safety label and classify by difference, improving calibration, reducing over-defense, and enabling token-level risk localization with 60.7% prompt rewriting success.

Gert Lek, Abele Mălan, Chaoyi Zhu, Pin-Yu Chen and 2 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

17/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 17 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 2/5

91%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightMicrosoft AI Microsoft RL for LLMs

Knowing When to Ask: Segment-Level Credit Assignment for LLM Tool Use

CARL assigns segment-level reinforcement learning credit at tool-use boundaries to teach models when external tools are needed, improving accuracy by up to 9.7 points while cutting unnecessary calls by 53%.

Abhijit Kumar, Zoey WU, Mohit Suley

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

17/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 17 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 2/5

91%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightSUN YAT-SEN UNIVERSITY Shenzhen Shenzhen Institutes of Advanced Alibaba Group Shenzhen MSU-BIT Segmentation

SIGMA: Semantic-Difference Instruction-Grounding Mask Annotator for Text-Driven Image Manipulation Localization

SIGMA uses semantic feature differencing with instruction-guided spatial priors to generate manipulation masks from edited images, producing a 1.1M training set that improves detectors by +18.34% F1.

Peiyu Zhuang, Jianquan Yang, Haodong Li, Zhuoying Cai and 5 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

17/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 17 of 20 reviewers recommend it

lenient 5/5

medium 9/10

strict 3/5

89%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Melbourne U Wisconsin-Madison AI oversight & deception

Combating Data Laundering in LLM Training

Data laundering transforms proprietary data to hide LLM training traces, and Synthesis Data Reversion restores detection by synthesizing likely transformed queries via goal-detail abstraction.

Muxing Li, Zesheng Ye, Sharon Li, Feng Liu

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

16/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 16 of 20 reviewers recommend it

lenient 5/5

medium 9/10

strict 2/5

89%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightGoogle DeepMind U California, San Diego U Waterloo Stanford; Noeri AI Tencent LLM agents & planning

Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction

Direct corpus interaction uses terminal tools to search raw corpora directly, bypassing fixed retrieval interfaces and substantially outperforming sparse, dense, and reranking baselines on agentic search benchmarks.

Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu and 14 more

Atlanta Poster Session 3, Thu, Dec 10, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 125 on Hugging Face · Code ★ 408

– ReadersNo votes yet

16/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 16 of 20 reviewers recommend it

lenient 5/5

medium 9/10

strict 2/5

89%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightUK AI Security Institute Anthropic Alan Turing Institute U Oxford The Alan Turing Institute Backdoors & data poisoning

Poisoning Attacks on LLMs Require a Near-constant Number of Poison Samples

Poisoning LLM pretraining requires only ~250 malicious documents regardless of dataset or model scale, revealing constant-cost backdoor injection risks for large models.

Alexandra Souly, Javier Rando, Ed Chapman, Xander Davies and 9 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 5 on Hugging Face

– ReadersNo votes yet

16/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 16 of 20 reviewers recommend it

lenient 5/5

medium 7/10

strict 4/5

89%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightNanyang Technological Shanghai Jiaotong Microsoft Shanghai Jiao Tong College of Computing and Data Sc Pose & motion

The Surprising Effectiveness of Video Diffusion Models for Hand Motion Reconstruction

ViDiHand leverages pretrained video diffusion models to reconstruct 4D hand poses directly from full egocentric video without detectors, substantially outperforming prior methods on ARCTIC, HOT3D, and HOI4D.

Yuxi Wang, Chengkai Jin, Yufei Liu, Wenqi Ouyang and 5 more

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 9 on Hugging Face · Code ★ 187

– ReadersNo votes yet

16/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 16 of 20 reviewers recommend it

lenient 5/5

medium 8/10

strict 3/5

89%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Cambridge Diffusion models

Recursively Trained Diffusion Models: Limiting Collapse Distribution and Spectral Characterization

Recursive diffusion training converges geometrically to a unique Gaussian-smoothed mixture limit via early-stopping drift, attenuating high-order spectral modes, with annealed truncation schedules asymptotically preventing collapse.

Nail B Khelifa, Richard Turner, Ramji Venkataramanan

Paris Poster Session 3, Thu, Dec 10, 12:30 PM–2:30 PM, Paris Poster Hall · Published 2026

– ReadersNo votes yet

16/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 16 of 20 reviewers recommend it

lenient 3/5

medium 9/10

strict 4/5

89%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightAnhui Polytechnic SUN YAT-SEN UNIVERSITY Nanyang Technological Nanyang Technology University, S Vision-language models

Text as Partial Constraint: Core–Residual Alignment for Robust Vision–Language Learning

TPC treats captions as partial constraints, aligning vision-language representations to a consensus semantic core while penalizing dependence on unsaid residuals, yielding robust zero-shot recognition and improved LVLM grounding.

Chengzhen Yu, Canran Xiao, SiYuan Ma, Yang Liu

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

16/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 16 of 20 reviewers recommend it

lenient 5/5

medium 9/10

strict 2/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Chicago Princeton Web & GUI agents

Mecha-nudges for Machines

Mecha-nudging changes online choice environments to systematically influence AI agents without harming human usability, and Etsy listings show a 0.143-bit rise in machine-usable information after ChatGPT's release.

Giulio Frey, Kawin Ethayarajh

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 7/10

strict 3/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightThe Cyprus Institute NCSR Demokritos U Oxford U Athens & Archimedes AI, At Cyprus Institute Sparse autoencoders

fmxcoders: Factorized Masked Crosscoders for Cross-Layer Feature Discovery

Standard crosscoders learn layer-localized features; fmxcoders use factorized weights and layer masking to recover cross-layer features, improving coherence and reconstruction across four LLMs.

Andreas D Demou, Panagiotis Koromilas, James Oldfield, Yannis Panagakis and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 3/5

medium 9/10

strict 3/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Cambridge Alignment & preference optimization

Don't Lose Focus: Activation Steering via Key-Orthogonal Projections

SKOP constrains harmful attention rerouting by preserving focus-token attention during steering, reducing utility degradation 5-7x at over 95% steering efficacy.

Haoyan Luo, Mateo Espinosa Zarlenga, Mateja Jamnik

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 10/10

strict 0/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightCzech Technical University of Pr Czech Technical Univeresity in P Czech Technical Univeresity in P Czech Techical Datasets & benchmarks

Doomed to Re-Annotate, Forever: The ImageNet Story

ReImageNet reveals ~12% of ImageNet-1k labels are wrong and reannotation boosts top-1 accuracy up to 1.2% for supervised models and 5, 6% for MLLMs.

Illia Volkov, Nikita Kisel, Tetiana Mishkina, Klara Janouskova and 1 more

Paris Poster Session 2, Wed, Dec 9, 5:00 PM–7:00 PM, Paris Poster Hall · Published 2026 · ▲ 3 on Hugging Face

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 4/5

medium 8/10

strict 3/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightCenter of Computational Neurosci New York Harvard Google DeepMind Training dynamics

Two Speeds of Learning: A Representation-Readout Decomposition of Grokking and Double Descent

A representation-readout decomposition attributes grokking and double descent to competing encoder and classifier dynamics, showing delayed generalization stems from gradual representation learning rather than lazy-to-rich transitions.

Chi-Ning Chou, Oscar Uzdelewicz, Neng-Chun Chiu, Yao-Yuan Yang and 1 more

Atlanta Poster Session 1, Wed, Dec 9, 10:00 AM–1:00 PM, Hall C1 · Published 2026

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 4/5

medium 9/10

strict 2/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Bristol Mila/École de technologie supéri Microsoft Research McGill U Calgary Probing & representation analysis

The Invisible Hand of Physics: When Video Diffusion Models Know More Than They Show

Video diffusion transformers encode physical plausibility linearly in latent states around 81% accuracy despite lacking predictive training, emerging inside the denoiser rather than the VAE.

Parsa Esmati, Somjit Nath, Katja Hofmann, Derek Nowrouzezahrai and 2 more

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 8/10

strict 2/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightPeking Microsoft Research, New England LLM evaluation & benchmarks

Calibration without Ground Truth

A label-free framework improves miscalibrated models using better-calibrated references, guaranteeing strict loss reduction without ground-truth labels via Bregman projection.

Yuqing Kong, Mingyu Song, Yizhou Wang, Yifan Wu

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 7/10

strict 3/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightThe Hong Kong Polytechnic The Hong Kong University of Scie U California, Berkeley Hong Kong Polytechnic ByteDance Inc.LLM agents & planning

VisInteract: Towards Dynamic Interactive Text-to-Visualization under Imperfect Queries

VisInteract introduces interactive text-to-visualization with imperfect queries via VisInteract-Bench and Vis-MCTS, boosting success by over 13% versus interactive baselines.

Wenxin XU, Jinwei Lu, Hwanhee Kim, Chen J Zhang and 3 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 7/10

strict 3/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightUC San Diego Shanghai Jiaotong U Science and Technology of Chin JD.com Zhejiang Video-language models

AdaCodec: A Predictive Visual Code for Video MLLMs

AdaCodec uses predictive visual codes to send full reference frames only when unpredictable, cutting video MLLM tokens by 7x while improving long-video benchmark scores and reducing latency.

Haowen Hou, Zhen Huang, Zheming Liang, Qingyi Si and 7 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026 · ▲ 7 on Hugging Face

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 4/5

medium 9/10

strict 2/5

88%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightNanyang Technological HKUST The Hong Kong University of Scie Segmentation

AdapToPASS: Ambiguity-aware Adaptive Spherical Transformer for Panoramic Semantic Segmentation

AdapToPASS is a bio-inspired spherical transformer that adaptively models contextual and geometric ambiguities to achieve robust panoramic semantic segmentation, outperforming state-of-the-art methods by up to 18.77% relative mIoU under unseen spherical transformations.

Soumyaratna Debnath, weiming zhang, Shriram Damodaran, Dingwen Xiao and 1 more

Sydney Poster Session 1, Tue, Dec 8, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 8/10

strict 2/5

88%Must read

NeurIPS 2026SpotlightTechnion Technion - Israel Institute of T Electrical Engineering Departmen Nvidia INRIA Unified multimodal models

MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image

MulTaBench benchmarks 40 multimodal tabular datasets and shows target-aware tuning of text and image embeddings improves predictive performance over frozen embeddings.

Alan Arazi, Eilam Shapira, Shoham Grunblat, Mor Ventura and 7 more

Paris Poster Session 5, Fri, Dec 11, 11:30 AM–1:30 PM, Paris Poster Hall · Published 2026 · ▲ 142 on Hugging Face

– ReadersNo votes yet

15/20 AI panelreviewers recommend it

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 15 of 20 reviewers recommend it

lenient 5/5

medium 8/10

strict 2/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightJohns Hopkins Massachusetts Institute of Techn Harvard the Kempner Institute for the St Google Deepmind / Harvard Pose & motion

MoSE3: Learning World-Space SE(3) at Every Pixel

MoSE3 predicts dense per-pixel world-space SE(3) motion from monocular video via point tracks and rigidity embeddings, achieving state-of-the-art 6-DoF estimation and 3D tracking.

Jiahuan Cheng, Zhiyi Li, Tian Xia, Ruojin Cai and 2 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 1 on Hugging Face

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 4/5

medium 7/10

strict 3/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightCMU, Carnegie Mellon Carnegie Mellon 3D vision & reconstruction

Point4D: Long-range 4D Motion Reconstruction

Point4D infers dense 3D point trajectories across multi-hundred-frame videos via a decoupled query-based motion decoder, outperforming prior short-window feed-forward 4D methods.

Minsik Jeon, Jay Karhade, Deva Ramanan, Shubham Tulsiani

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 4/5

medium 9/10

strict 1/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightFudan Hong Kong University of Science Mila / Université de Montréal City University of Hong Kong Hong Kong University of Science Deep RL

Local Guidance, Global Impact: Gaussian-Reshaped Trust Region Unlocks Behavior Transitions

Gaussian trust region reshaping replaces monotonic divergence penalties with bounded non-monotonic constraints, unlocking efficient behavior transitions in non-stationary reinforcement learning.

Bingxu Liu, Jiashun Liu, Johan Obando Ceron, Hao Wang and 4 more

Sydney Poster Session 6, Thu, Dec 10, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 4/5

medium 9/10

strict 1/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightNational Taiwan University / CyC National Taiwan University Appier National Taiwan Jailbreaks & red teaming

Expected Harm: Rethinking Safety Evaluation of (Mis)Aligned LLMs

Expected Harm weights jailbreak severity by execution likelihood to reveal inverse risk calibration, where models over-refuse costly threats yet remain vulnerable to cheap, high-likelihood attacks that double jailbreak success.

Yen-Shan Chen, Zhi Rui Tam, Cheng-Kuang Wu, Yun-Nung (Vivian) Chen

Sydney Poster Session 5, Thu, Dec 10, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026 · ▲ 21 on Hugging Face

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 5/5

medium 7/10

strict 2/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU California, Berkeley UC Berkeley, Transluce AI Multi-agent LLM systems

Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems

Simple norm enforcement for AI agents is exploited for competitive gain, but mechanisms tracking reliability with escalating penalties resist exploitation across multi-agent environments.

Yaowen Ye, Jacob Steinhardt

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 5/5

medium 8/10

strict 1/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightGenesis Molecular AI MIT Massachusetts Institute of Techn Genesis Therapeutics Genesis Molecular AI Flow matching

Few-Step Cofolding with All-Atom Flow Maps

DeCAF distills all-atom biomolecular cofolding diffusion models into few-step flow maps with SE(3)-aligned endpoint losses, improving accuracy and physical validity at strict inference budgets.

Gianluca Scarpellini, Ron Shprints, Peter Holderrieth, Juno Nam and 6 more

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 3/5

medium 10/10

strict 1/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightEPFL U Michigan, Ann Arbor USI TU Wien, AITHYRA, University of KAIST LLM reasoning & chain of thought

MUX: Continuous Reasoning via Multiplexed Tokens

MUX compresses reasoning into continuous multiplexed tokens via lossless superposition, accelerating reasoning and outperforming latent baselines across 32 settings.

Ayhan Suleymanzade, Halil Alperen Gozeten, Michael Bronstein, Ismail Ilkan Ceylan and 1 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 4/5

medium 8/10

strict 2/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightSUN YAT-SEN UNIVERSITY Sun Yat-Sen Sun Yat-sen Zhejiang China University of Mining and T Time series classification & imputation

TSQAgent: Rating Time Series Data Quality via Dedicated Agentic Reasoning

TSQAgent uses collaborative agent roles and external analytical tools to automatically identify relevant time series quality dimensions and perform quantitative comparisons, substantially improving LLM assessment and downstream data selection.

Shunyu Wu, Dan Li, Haozheng Ye, Weibin Feng and 5 more

Sydney Poster Session 2, Tue, Dec 8, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 5/5

medium 8/10

strict 1/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightApple Alignment & preference optimization

Theoretical Limits of Language Model Alignment

Optimal alignment gains follow a Jeffreys-divergence limit, best-of-N approaches it, and reward hacking grows with error while ensembling mitigates it.

Lucas Monteiro Paes, Natalie Mackraz, Barry-John Theobald, Federico Danieli

Atlanta Poster Session 4, Thu, Dec 10, 4:30 PM–7:30 PM, Hall C1 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 4/5

medium 7/10

strict 3/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightCMU, Carnegie Mellon Department of Computer Science, U Washington Google U North Carolina at Chapel Hill Pose & motion

MolmoMotion: Forecasting Point Trajectories in 3D with Language Instruction

MolmoMotion predicts goal-conditioned 3D point trajectories from visual history and language, outperforming baselines on PointMotionBench and improving robot manipulation and video synthesis.

Jianing Zhang, Chenhao Zheng, Yajun Yang, Rustin Soraki and 6 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026 · ▲ 53 on Hugging Face · Code ★ 151

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 5/5

medium 7/10

strict 2/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightMcGill U Tennessee, Knoxville McGill & MBZUAI California Institute of Technolo MBZUAI Navigation & SLAM

GeoWind2Plan: Mission-Time 3D Urban Wind Prediction for Energy-Efficient UAV Planning

GeoWind2Plan predicts mission-time 3D urban wind via neural operators to enable energy-efficient UAV planning in seconds, reducing energy by up to 12.7% versus wind-agnostic paths.

Shaoxiang Qin, Xiongye Xiao, Yucheng Zhao, Fuyuan Lyu and 5 more

Atlanta Poster Session 5, Fri, Dec 11, 10:00 AM–1:00 PM, Hall C1 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 5/5

medium 6/10

strict 3/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightU Zurich Ecole Polytechnique Paris / Tech Graz University of Technology Molecules & drug discovery

Generating Physically Consistent Molecules with Energy-Based Models

EBMol learns atom-additive scalar potentials via flow-inspired restoring field matching to generate physically consistent 3D molecules with state-of-the-art results.

Christoph Griesbacher, Lea Bogensperger, Andreas Habring, Thomas Pock

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 3/5

medium 10/10

strict 1/5

86%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightMassachusetts Institute of Techn MIT Diffusion models

The Value of Covariance Matching in Gaussian DDPMs and the Lanczos Sampler

Matching full posterior covariance in Gaussian DDPMs reduces path-KL error to O(1/T²), and the matrix-free Lanczos Gaussian sampler achieves this with exponentially decaying approximation error using only Jacobian-vector products.

Sahil Akhtar, Aymane El Gadarri, Vivek Farias, Adam Jozefiak

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

14/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 14 of 20 reviewers recommend it

lenient 2/5

medium 9/10

strict 3/5

83%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightSoutheast Nanjing University of Informatio Ant Group Efficient inference & serving

TreeGraft: Adaptive Multi-Drafter Grafting for Tree-Based Speculative Decoding

TreeGraft combines small and large drafters with a scheduler to build shared draft trees, boosting speculative decoding by 15.1% over single-drafter methods.

Jiaming Fan, Daming Cao, Canchen Huang, Jiale Fu and 5 more

Sydney Poster Session 4, Wed, Dec 9, 5:00 PM–8:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

13/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 13 of 20 reviewers recommend it

lenient 4/5

medium 7/10

strict 2/5

83%Must read

?Must readVote to see the score

NeurIPS 2026SpotlightTechnion - Israel Institute of T Technion - Israel Institute of T IT University of Copenhagen Graph neural networks

Weisfeiler-Leman Is Incomplete on Simple Spectrum Graphs, so Canonicalize Them

For every k, k-WL fails to distinguish some non-isomorphic simple-spectrum graphs, so PRiSM provides the first complete canonicalization of their eigendecompositions to enable universal approximation.

Snir Hordan, Nadav Dym, Tim Seppelt

Sydney Poster Session 3, Wed, Dec 9, 10:00 AM–1:00 PM, Hall 1-4 · Published 2026

– ReadersNo votes yet

13/20 AI panelreviewers recommend it

Readers and the AI panel: vote on this paper to see what they said.

Only vote on papers you've read. Sign in with GitHub to vote.

AI panel: 13 of 20 reviewers recommend it

lenient 2/5

medium 8/10

strict 3/5

Show 40 more papers