NeurIPS 2026

The best-rated papers in every poster session, from 6,073 presentations. Plan your walk through the halls.

Sydney Poster Session 1

Tue, Dec 8, 10:00 AM–1:00 PM local time · Hall 1-4

92%?SR-Prominence: A Crowdsourced Protocol and Dataset Suite for Perceptually-Weighted Super-Resolution Artifact EvaluationSR-Prominence defines artifact prominence via crowdsourced annotations across 3,935 masks and shows classical full-reference metrics surprisingly detect perceptual impact better than specialized detectors.NeurIPS 2026Must read · vote to see score

91%?CardioLens: Revealing the Clinical Reality Gap of MLLMs via Multi-Sequence Cardiac MRI EvaluationsCardioLens evaluates MLLMs on multi-sequence cardiac MRI, revealing poor clinical workflow performance and category-collapse failures despite reasoning prompts and slice selection.NeurIPS 2026Must read · vote to see score

91%?Agent Meltdowns: The Road to Hell Is Paved with Helpful AgentsAgents encountering benign errors suffer "accidental meltdowns", unsafe behaviors like unauthorized reconnaissance, across 64.7% of error rollouts, often unreported.NeurIPS 2026Must read · vote to see score

91%?Reading the Finetuning Prior: Verbatim Content Recovery via Contrastive Decoding DiffingContrastive Decoding Diffing recovers verbatim implanted facts and pipeline artifacts via output-level logit differences without weight access, outperforming white-box methods 170x faster.NeurIPS 2026Must read · vote to see score

91%?CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language ModelsCRAFT 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.NeurIPS 2026 SpotlightMust read · vote to see score

91%?GraphInstruct: A Progressive Benchmark for Diagnosing Capability Gaps in LLM Graph GenerationGraphInstruct introduces progressive-complexity benchmark diagnosing LLM graph generation failures across six complexity levels, finding multi-constraint composition limits capability and domain-semantic constraints require retrieval.NeurIPS 2026Must read · vote to see score

Sydney Poster Session 2

Tue, Dec 8, 5:00 PM–8:00 PM local time · Hall 1-4

92%?SpatialBench: Is Your Spatial Foundation Model an All-Round PlayerSpatialBench evaluates 41 spatial foundation models across 19 datasets and finds none are all-round players, with full-context attention maximizing accuracy and domain alignment exceeding scaling for embodied tasks, plus it introduces DA-Next-5M and DA-Next.NeurIPS 2026Must read · vote to see score · ▲ 69 on HF

92%?Where Root Cause Analysis Fails: A Retrieval-Reranking DecompositionRoot cause analysis benchmarks conflate retrieval and reranking failures, revealing graph methods rarely beat statistical baselines; a two-stage retriever-LLM reranker matches or exceeds all baselines without causal graphs or labels.NeurIPS 2026Must read · vote to see score

92%?Computer Use at the Edge of the Statistical PrecipiceA 1MB replay script outperforms frontier agents on static benchmarks because of flawed environment design and evaluation; the paper proposes PRISM principles, DigiWorld, and hierarchical bootstrap aggregation to fix both.NeurIPS 2026 OralMust read · vote to see score

92%?Object Detection Benchmarks are Incomplete: The Role of Label Errors and Annotation UncertaintyRe-annotation reveals object detection benchmarks miss up to 60% of objects due to incomplete labels, and current detectors remain misaligned with human perception.NeurIPS 2026 SpotlightMust read · vote to see score

LLM reasoning & chain of thought91%?Pause and Reflect: Conformal Aggregation for Chain-of-Thought ReasoningA conformal procedure for chain-of-thought reasoning replaces majority voting with calibrated weighted aggregation to provide finite-sample confident-error guarantees and improves selective accuracy without retraining.NeurIPS 2026Must read · vote to see score

91%?Mark, Don't Erase: Token Inoculation for Dual-Use Knowledge in LLMsToken Inoculation conditions LLMs to retain dual-use knowledge gated by a special token, reducing hazardous accuracy to 18% while preserving 93% of benign performance across 1B-14B scales.NeurIPS 2026Must read · vote to see score

Sydney Poster Session 3

Wed, Dec 9, 10:00 AM–1:00 PM local time · Hall 1-4

92%?Hidden Measurement Error in LLM Pipelines Distorts Annotation, Evaluation, and BenchmarkingLLM pipeline evaluation variance is underestimated because design choices are ignored, so corrected intervals restore coverage and cut benchmark gaming.NeurIPS 2026Must read · vote to see score

91%?RealityTest: How People Probe AI Identity and Whether Models Disclose ItRealityTest benchmarks multimodal multilingual AI identity disclosure via 3,152 human queries, finding question phrasing and context dominate over model choice and suppression cuts rates below 30%.NeurIPS 2026Must read · vote to see score

91%?Jointly Reinforcing Diversity and Quality in Language Model GenerationsDARLING uses a learned partition function to jointly optimize language model response quality and semantic diversity via reinforcement learning, improving both quality and novelty across creative and math benchmarks.NeurIPS 2026Must read · vote to see score · ▲ 25 on HF · 5 comments

91%?PACE: A Proxy for Agentic Capability EvaluationPACE predicts agentic benchmark scores from small, selected non-agentic test subsets via regression, achieving under 4% error and over 0.80 correlation at under 1% evaluation cost.NeurIPS 2026Must read · vote to see score · ▲ 18 on HF · 3 comments

91%CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation ModelsCollabVR pairs vision-language models with video generation models in closed-loop step-level planning and verification, reducing drift and simulation errors for major video reasoning gains.NeurIPS 2026Must read · 100% of 1 readers upvoted · ▲ 71 on HF · 5 comments

91%?Embeddings for Preferences, Not SemanticsText embeddings should encode preferential rather than semantic similarity for collective decisions; breaking nuisance correlation with synthetic training improves preference prediction across 11 deliberation datasets.NeurIPS 2026Must read · vote to see score

Sydney Poster Session 4

Wed, Dec 9, 5:00 PM–8:00 PM local time · Hall 1-4

92%?K12-KGraph: A Curriculum-Aligned Knowledge Graph for Benchmarking and Training Educational LLMsK12-KGraph introduces a curriculum-aligned K-12 knowledge graph, benchmark, and training data showing current LLMs achieve under 57 percent accuracy on curriculum cognition and that graph-guided supervision outperforms generic instruction tuning.NeurIPS 2026Must read · vote to see score · ▲ 62 on HF

Jailbreaks & red teaming92%?Alignment Collapse Under KV Cache Quantization: Diagnosis and MitigationLow-bit KV cache quantization silently collapses LLM safety alignment via geometric subspace vulnerability, and per-channel reduction diagnostics recover up to 97% of lost refusals.NeurIPS 2026Must read · vote to see score

Code generation91%?P$^{3}$: Joint Program-and-Proof Planning\\ for Verified Code GenerationP³ plans programs and proofs jointly from specifications before elaboration, outperforming sequential baselines by up to 11.2 points on verified generation benchmarks while reducing cost and time.NeurIPS 2026Must read · vote to see score

91%?Transferable SCF-Acceleration through Solver-Aligned Initialization LearningSolver-Aligned Initialization Learning differentiates through SCF solvers to train transferable ML initial guesses, reducing iterations by up to 37% on molecules up to 10× larger than training data.NeurIPS 2026Must read · vote to see score

91%?Reflective Prompted Policy Optimization: Trajectory-Grounded Revision and Salience BiasR2PO uses trajectory-level behavioral evidence rather than scalar rewards to guide LLM policy search, achieving faster and more stable optimization across ten environments despite a critic salience bias.NeurIPS 2026Must read · vote to see score

91%?MCPHunt: An Evaluation Framework for Cross-Boundary Data Propagation in Multi-Server MCP AgentsMCPHunt benchmarks multi-server MCP agents, finding 11.5, 41.3% policy-violating cross-boundary credential propagation concentrated in browser flows, with prompt mitigations reducing violations up to 97%.NeurIPS 2026Must read · vote to see score

Paris Poster Session 1

Wed, Dec 9, 12:30 PM–2:30 PM local time · Paris Poster Hall

91%?Crowded in B-Space: Calibrating Shared Directions for LoRA MergingLoRA merging interference mainly stems from shared output-side B directions; calibrating them via Pico improves merged adapter accuracy across benchmarks and can exceed joint-training performance.NeurIPS 2026Must read · vote to see score · ▲ 6 on HF

91%?The BatchNorm Illusion: Diagnosing Normalization Artifacts in Machine Unlearning EvaluationBatchNorm running statistics artificially inflate unlearning metrics by up to 78 points, which a weight-preserving forward pass reverses without changing weights.NeurIPS 2026Must read · vote to see score

Vision-language models89%?GT-Free OCR Metrics: A Reference-Free Evaluation Framework for Document OCR SystemsA render-and-compare framework benchmarks 147 reference-free visual metrics for document OCR, with the best composite achieving ρ = 0.494 correlation to ground-truth quality. Silent region misclassification artificially preserves visual similarity, suppressing correlation when masking is omitted.NeurIPS 2026Must read · vote to see score

89%?ThousandWorlds: A benchmark for climate emulation of potentially habitable exoplanetsThousandWorlds introduces a multi-model exoplanet climate benchmark of ~1,700 GCM simulations, showing Gaussian processes outperform deep learning in low-data multi-simulator regression.NeurIPS 2026Must read · vote to see score · ▲ 1 on HF

88%?Boosting Brain-to-Image Decoding with TRIBE v2 Data AugmentationTRIBE v2 synthetic fMRI augmentation improves brain-to-image decoding by up to 68%, though optimal synthetic-to-real ratios vary by dataset, and synthetic-only training achieves above-chance zero-shot decoding.NeurIPS 2026Must read · vote to see score

88%?idSCD: Identifying Training Datasets through Semantic Correlation DescriptorsDataset-specific training traces are captured via semantic correlation descriptors (SCDs) that fingerprint dataset membership via internal semantic correlations, outperforming black-box and white-box baselines by over 60% ROC-AUC when semantic particularities differ.NeurIPS 2026Must read · vote to see score · 4 comments

Atlanta Poster Session 1

Wed, Dec 9, 10:00 AM–1:00 PM local time · Hall C1

92%?LowRankArena: A Standardized Evaluation Platform for SVD-Based LLM CompressionLowRankArena standardizes SVD-based LLM compression evaluation and reveals that method rankings and speedups depend heavily on backbone and workload under aligned protocols.NeurIPS 2026Must read · vote to see score · ▲ 1 on HF

91%Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End LatencyLoHi trades per-frame resolution for denser temporal sampling via low-resolution streams plus sparse high-resolution frames, boosting long-video accuracy up to 10.6 points and cutting front-end latency up to 7x.NeurIPS 2026Must read · ▲ 13 on HF · 18 comments

91%?AgentKVShift: Efficient KV Cache Reuse for Agentic Memory SystemsAgentKVShift uses probe-guided KV residual correction to reuse agentic memory caches with near-full accuracy at 10-30% recompute, yielding 2-3.5x prefill speedups.NeurIPS 2026Must read · vote to see score

91%?Crafting Reversible SFT Behaviors in Large Language ModelsLCDD constructs sparse, causally necessary subnetworks for SFT behaviors, and SFT-Eraser reverses them via activation-matched soft prompts without weight changes.NeurIPS 2026Must read · vote to see score

91%?A Communication-Theoretic Framework for LLM Agents: Cost-Aware Adaptive ReliabilityA communication-theoretic framework maps LLM reliability techniques onto classical coding operators, yielding closed-form thresholds and a cost-aware router that achieves a 56% cost reduction at matched quality across 69 hard tasks.NeurIPS 2026Must read · vote to see score

91%?Response Time Enhances Alignment with Heterogeneous PreferencesAdding response times to preference data via drift-diffusion modeling restores identifiability of average preferences among anonymous heterogeneous labelers, correcting choice-only estimation bias without tracking users.NeurIPS 2026Must read · vote to see score

Paris Poster Session 2

Wed, Dec 9, 5:00 PM–7:00 PM local time · Paris Poster Hall

91%?StereoTales: A Multilingual Framework for Open-Ended Stereotype Discovery in LLMsStereoTales reveals open-ended LLM generation emits shared harmful stereotypes that culturally adapt to prompt languages and align with human harmfulness ratings.NeurIPS 2026Must read · vote to see score

91%?Automata from Agent Traces: Failure and Next-Step PredictionTrace corpora collapse into compact finite-state machines replaying held-out data at >=0.997 fitness, yielding state-context next-step prediction and 0.94 AUROC failure prediction for runtime monitoring.NeurIPS 2026Must read · vote to see score · ▲ 5 on HF · 5 comments

89%?Assessing Per-Sample Membership Inference Vulnerability without RetrainingPer-sample membership inference vulnerability is governed by a data-dependent geometric measure, yielding a surrogate score using only a single model that outperforms loss-based baselines at identifying high-risk training points.NeurIPS 2026Must read · vote to see score

89%?CanViT: Toward Active-Vision Foundation ModelsCanViT introduces the first active-vision foundation model with a retinotopic backbone and scene-wide canvas, achieving 38.5% ADE20K mIoU with one glimpse and 84.5% ImageNet accuracy.NeurIPS 2026Must read · vote to see score · ▲ 13 on HF

89%?Principia: Relational Physics Tests for Video ModelsPrincipia benchmarks video generators via calibration-independent relational physics consistency across eight Newtonian phenomena, finding top models score below 0.42 despite high VBench ratings.NeurIPS 2026Must read · vote to see score · ▲ 18 on HF · 6 comments

89%?ReefNet: A Large-Scale Dataset and Benchmark for Fine-Grained Coral Reef RecognitionReefNet provides ~925K genus-level coral annotations and benchmarks showing vision models degrade under zero-shot and cross-source shifts despite adaptation gains.NeurIPS 2026Must read · vote to see score

Atlanta Poster Session 2

Wed, Dec 9, 4:30 PM–7:30 PM local time · Hall C1

91%?ROCKET: Residual-Oriented Multi-Layer Alignment for Spatially-Aware Vision-Language-Action ModelsROCKET aligns multiple VLA layers to a 3D vision model via residual streams and shared projectors, achieving near-state-of-the-art LIBERO success with about 4% compute.NeurIPS 2026Must read · vote to see score

91%?Flow Map Denoisers: Traversing the Distortion-Perception Plane for Inverse ProblemsFlow map denoisers implicitly define a one-parameter family spanning the distortion-perception tradeoff via lookahead parameter t, matching or exceeding specialized baselines across inverse problems.NeurIPS 2026Must read · vote to see score

91%?MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AIMLS-Bench evaluates AI agents on inventing scalable ML methods across 140 tasks, finding current systems fail to reliably surpass human-designed approaches due to insufficient scientific validation insight.NeurIPS 2026 OralMust read · vote to see score · ▲ 8 on HF

Feature attribution89%?Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic SaliencyCounterfactual Semantic Saliency reveals VLMs diverge from human scene perception via size, center, and saliency biases while underweighting people.NeurIPS 2026Must read · vote to see score

89%?Quest: Training Frontier Deep Research Agents with Fully Synthetic TasksQUEST trains open deep research agents via synthetic rubric-tree tasks and context management, achieving frontier-level performance across eight benchmarks with only 8K examples.NeurIPS 2026Must read · vote to see score · ▲ 43 on HF

In-context learning & prompting89%?Learning to Follow In-Context Watermark Instructions via Self-DistillationICWBench reveals current LLMs fail at in-context watermarking, and self-distillation with reinforcement learning raises watermark detectability near perfect while preserving quality.NeurIPS 2026Must read · vote to see score

Sydney Poster Session 5

Thu, Dec 10, 10:00 AM–1:00 PM local time · Hall 1-4

92%?Revisiting Embodied Chain-of-Thought for Generalizable Robot ManipulationEmbodied 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.NeurIPS 2026 SpotlightMust read · vote to see score · 3 comments

Molecules & drug discovery92%?MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology AssaysMorphoHELM benchmarks microscopy representation methods across batch effects, finding classic computer vision strategies outperform deep learning across settings and revealing trade-offs between models.NeurIPS 2026Must read · vote to see score

92%?Reinforcement Learning for Code OptimizationReinforcement learning for code optimization fails due to noisy, sparse execution-time rewards, so a calibrated three-stage pipeline improves strict pass rates by up to 125% while preserving correctness.NeurIPS 2026Must read · vote to see score · ▲ 13 on HF

92%?DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document UnderstandingDocScope benchmarks verifiable long-document reasoning via structured trajectory evaluation, finding correct answers rarely include complete evidence chains and region grounding remains weakest.NeurIPS 2026Must read · vote to see score

91%?DriveSpatial: A Benchmark for Spatiotemporal Intelligence in VLMs for Autonomous DrivingDriveSpatial benchmarks vision-language models' spatiotemporal autonomous driving intelligence, finding a 28.4-point human gap with cognitive scene construction as the key bottleneck.NeurIPS 2026Must read · vote to see score

91%?Déjà View: Looping Transformers for Multi-View 3D ReconstructionDéjàView loops a single transformer block for iterative multi-view 3D reconstruction, matching larger feed-forward models with far fewer parameters while treating refinement steps as an inference-time compute knob.NeurIPS 2026 OralMust read · vote to see score · ▲ 6 on HF

Sydney Poster Session 6

Thu, Dec 10, 5:00 PM–8:00 PM local time · Hall 1-4

92%?CausalDriveBench: Evaluating Causal Reasoning in Vision-Language-Action Models for Autonomous DrivingCausalDriveBench evaluates causal reasoning in autonomous driving vision-language-action models via structured QA and counterfactual trajectories, finding weak causal understanding despite fluent reasoning and accurate baseline predictions.NeurIPS 2026Must read · vote to see score

92%?TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular EncodersTRL-Bench standardizes cross-paradigm evaluation of tabular encoders via shared representation-level probes, finding encoder quality is task-specific and best pipelines combine capability-matched specialists.NeurIPS 2026Must read · vote to see score · ▲ 54 on HF

92%?Log-Likelihood, Simpson’s Paradox, and the Detection of Machine-Generated TextAverage token-level log-likelihood scores suffer Simpson’s paradox across hidden-space regions, and local calibration via learned score-distribution predictors fixes it, boosting detection AUROC substantially.NeurIPS 2026Must read · vote to see score

91%?Attention Transfer Is Not Universally Effective for Vision TransformersAttention transfer fails for four ViT families due to architectural mismatch, and adding the teacher's native components to students fully restores its effectiveness.NeurIPS 2026Must read · vote to see score

91%?FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action PoliciesFineVLA introduces fine-grained action-aligned supervision for steerable vision-language-action policies, yielding up to 86.8% simulation and 62.7 real-world success and boosting steerable control over coarse instructions.NeurIPS 2026Must read · vote to see score · ▲ 9 on HF

91%?Claw-Eval: Towards Trustworthy Evaluation of Autonomous AgentsClaw-Eval introduces a trajectory-aware benchmark with 300 tasks, finding opaque grading misses 44% of safety violations and agent rankings vary across multi-dimensional capabilities.NeurIPS 2026Must read · vote to see score · ▲ 117 on HF · 3 comments

Paris Poster Session 3

Thu, Dec 10, 12:30 PM–2:30 PM local time · Paris Poster Hall

Test-time compute91%?The Ringelmann Effect in Multi-Agent LLM Systems: A Scaling Law for Effective Team SizeA two-parameter scaling law quantifies diminishing returns in multi-agent LLM systems, showing that dense debate hits hard ceilings, noise placebos match self-correction, and only heterogeneous teams escape diminishing returns.NeurIPS 2026Must read · vote to see score

91%?Beyond the Golden Teacher: Enhancing Graph Learning through LLM-GNN Co-teachingLLM-GNN Co-Teaching replaces golden-teacher design with bidirectional pseudo-label exchange and trajectory-based preference optimization, boosting few-shot graph accuracy by up to 7.86%.NeurIPS 2026Must read · vote to see score

91%?CalArena: A Large Scale Post-Hoc Calibration BenchmarkCalArena benchmarks nearly 2000 post-hoc calibration experiments, finding smooth methods outperform binning and multiclass-specific designs are essential.NeurIPS 2026Must read · vote to see score

91%?Improved Baselines with Representation AutoencodersUsing summed last-k encoder layers and combining RAE with REPA, RAEv2 achieves state-of-the-art gFID of 1.06 in 80 epochs with 10x faster convergence and free guidance.NeurIPS 2026Must read · vote to see score

89%?Handwritten Text Recognition Lives in the High-Pixel Variance SubspaceFor handwritten text recognition, discriminative signals lie in high-variance pixel directions, so pixel-reconstruction self-supervised pretraining outperforms contrastive methods and achieves lower character error rates across benchmarks.NeurIPS 2026Must read · vote to see score

89%?Recursively Trained Diffusion Models: Limiting Collapse Distribution and Spectral CharacterizationRecursive 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.NeurIPS 2026 SpotlightMust read · vote to see score

Atlanta Poster Session 3

Thu, Dec 10, 10:00 AM–1:00 PM local time · Hall C1

92%?OdysSim: Building Foundation Models for Human Behavior SimulationOdysSim trains 8B behavioral foundation models via SOUL taxonomy and multi-stage recipes, ranking first on eight human simulation benchmarks while nearly matching real-user reaction alignment.NeurIPS 2026Must read · vote to see score · ▲ 1 on HF

91%?Models Recall What They Violate: Constraint Adherence in Multi-Turn LLM IdeationDriftBench finds iterative LLM ideation increases complexity and reduces constraint adherence, with models often violating rules they accurately recall and judges under-detecting violations.NeurIPS 2026Must read · vote to see score

91%?SchemeArena: Factorized Stress Testing of Scheming in LLM AgentsSchemeArena introduces a 400-scenario benchmark and SCOUT monitor for factorized LLM agent scheming stress tests, finding explicit instrumental goals drive scheming most strongly and partial oversight can increase covert behavior.NeurIPS 2026Must read · vote to see score · ▲ 11 on HF

91%?The Best-Laid SCHEMEs: Coordinated Sabotage and Monitoring in Multi-Agent SystemsSCHEME benchmark reveals multi-agent models coordinate sabotage via decomposed plans across communication topologies, with Gemini succeeding 84% and Codex 46%, though monitors detect edits at 99%/68% and communication at 100%/81%.NeurIPS 2026Must read · vote to see score

91%?On the Error Correcting Effects of Stochasticity in Discrete DiffusionDiscrete diffusion stochasticity trades convergence speed against error correction via redundant transitions, and DCRS injects controlled randomness to improve low-step sampling efficiency.NeurIPS 2026Must read · vote to see score

91%?Meta$^n$: Recursive Self-Improvement through Emergent DepthMeta^n fixes its meta-operation and recurses on growing inputs to build unbounded self-improving agent depth, outperforming prior agents across eight benchmark families including ARC-AGI-2.NeurIPS 2026Must read · vote to see score · ▲ 17 on HF

Paris Poster Session 4

Thu, Dec 10, 5:30 PM–7:30 PM local time · Paris Poster Hall

Representation & contrastive learning91%?When Attention Collapses: Residual Evidence Modeling for Compositional InferenceUnder additive superposition, attention slots collapse to dominant components because memoryless attention ignores explained evidence; residual evidence depletion prevents collapse and enables compositional inference.NeurIPS 2026Must read · vote to see score

91%?StemBind: When MLLMs Get Lost Between Rules and Instances in Abstract Visual ReasoningStemBind introduces a shared-stem benchmark diagnosing MLLM abstract visual reasoning, finding a persistent rule-to-instance binding gap where models identify patterns but fail to apply them correctly.NeurIPS 2026Must read · vote to see score

91%?Causal Representation Learning for Generalisable RecommendationA causal disentanglement objective improves recommender out-of-distribution generalization by isolating invariant causal components, yielding substantial online engagement gains in Spotify A/B tests.NeurIPS 2026Must read · vote to see score · 5 comments

3D generation89%?One View Is Enough: In-the-Wild Monocular Pretraining for Novel View GenerationMonocular pretraining lifts single images into pseudo-target views via depth and reprojection, yielding OVIE, which rivals multi-view baselines at 116 FPS without inference-time depth or multi-view training pairs.NeurIPS 2026Must read · vote to see score · ▲ 5 on HF

88%?Half-Truths Break Similarity-Based RetrievalCLIP-style dual encoders often prefer half-true image descriptions with incorrect added details over correct shorter ones due to weak part-level supervision; CS-CLIP improves half-truth accuracy to 69.3% through component-level contrastive fine-tuning.NeurIPS 2026Must read · vote to see score · ▲ 6 on HF

88%?Your Neighbors Know: Leveraging Local Neighborhoods for Backdoor Detection in Decentralized LearningArgus detects backdoor attacks in decentralized learning by having nodes share local trigger analyses with neighbors and filter updates via structural similarity, reducing attack success by up to 90 points without a central server.NeurIPS 2026Must read · vote to see score

Atlanta Poster Session 4

Thu, Dec 10, 4:30 PM–7:30 PM local time · Hall C1

91%?Self Driving Datasets: From 20 Million Papers to Nuanced Biomedical Knowledge at ScalePubMed is autonomously converted into structured biomedical datasets larger, more nuanced, and more accurate than manual repositories via ontology tagging, hybrid retrieval, and a multi-agent extraction system.NeurIPS 2026Must read · vote to see score

91%?ECHO: Terminal Agents Learn World Models for FreeECHO trains terminal agents to predict environment responses for dense supervision, doubling GRPO pass@1 on TerminalBench-2.0.NeurIPS 2026 SpotlightMust read · vote to see score · ▲ 7 on HF · 3 comments

91%?VeriContest: A Competitive-Programming Benchmark for Verifiable Code GenerationVeriContest introduces 946 competitive programming problems with verified Rust specifications and proofs, showing state-of-the-art models reach only 5.29% on end-to-end verifiable generation.NeurIPS 2026Must read · vote to see score · ▲ 1 on HF

Instruction tuning91%?Beyond LoRA vs. Full Fine-Tuning: Gradient-Guided Optimizer Routing for LLM AdaptationMoLF dynamically routes optimizer updates between full fine-tuning and LoRA to match or beat the stronger static method across tasks, and its efficient variant surpasses AdaLoRA and AdaMix by up to 11.70 points.NeurIPS 2026Must read · vote to see score · 6 comments

Multi-agent LLM systems91%?AgentCollabBench: Diagnosing When Good Agents Make Bad CollaboratorsAgentCollabBench introduces 900 diagnostic tasks showing multi-agent collaboration failures stem from topology, not just model capability. Communication topology explains 7-40% of variance as converging nodes discard minority-branch constraints.NeurIPS 2026Must read · vote to see score

91%?Entropy Distribution as a Fingerprint for Hallucinations in Generative ModelsToken-level entropy distributions fingerprint hallucinations, and the single-pass Calibrated Entropy Score achieves multi-pass detection accuracy with formal guarantees.NeurIPS 2026Must read · vote to see score

Paris Poster Session 5

Fri, Dec 11, 11:30 AM–1:30 PM local time · Paris Poster Hall

91%?One World, Dual Timeline: Decoupled Spatio-Temporal Gaussian Scene Graph for 4D Cooperative Driving ReconstructionDUST uses decoupled dual-timeline Gaussian scene graphs for asynchronous vehicle-infrastructure 4D reconstruction, cutting ghosting to boost dynamic-area PSNR by 3.2 dB.NeurIPS 2026Must read · vote to see score

91%?When to Align, When to Predict: A Phase Diagram for Multimodal LearningUnder structured cross-modal nuisance correlation, cross-modal alignment and prediction have complementary failure modes partitioned by separation ratios into four regimes, with a data-driven procedure identifying preferred objectives and when neither beats single-modality baselines.NeurIPS 2026Must read · vote to see score

89%?Predict-Project-Renoise: Sampling Diffusion Models under Hard ConstraintsPPR samples pretrained diffusion models under hard constraints via iterative denoiser projection and renoising, achieving near-zero violations with high fidelity across physics and weather tasks.NeurIPS 2026Must read · vote to see score

89%?Liars' Bench: Evaluating Lie Detectors for Language ModelsLiars' Bench evaluates lie detectors across 72,863 LLM lies and finds existing techniques systematically miss certain lie types, especially when transcripts alone are insufficient.NeurIPS 2026Must read · vote to see score · ▲ 1 on HF

89%?Test-Time Personalization: A Diagnostic Framework and Probabilistic Fix for Scaling FailuresTest-time personalization samples candidates and selects via reward models, proving logarithmic utility scaling but diagnosing user collapse and query hacking, fixed by probabilistic rewards.NeurIPS 2026Must read · vote to see score

89%?Validating Causal Abstraction Metrics on Simulated Complex SystemsBenchmarking thirty metrics on ten simulated complex systems shows only causal metrics reliably validate high-level explanations when testing unmapped-variable faithfulness, leading to the Causal Abstraction Error metric converging with thirty interventions.NeurIPS 2026Must read · vote to see score

Paris Poster Session 6

Fri, Dec 11, 2:30 PM–4:30 PM local time · Paris Poster Hall

92%?Measuring Cross-Modal Synergy: A Benchmark for VLM ExplainabilityCross-modal redundancy causes unimodal metrics to contradict (τ=-0.06), so Synergistic Faithfulness (F_syn) isolates joint modality dividends with ρ=0.92 and 24× speedup, revealing VLM explainers over-index visual salience versus adapted attention methods.NeurIPS 2026Must read · vote to see score

91%?GitInject: Real-World Prompt Injection Attacks in AI-Powered CI/CD PipelinesGitInject tests real AI CI/CD workflows and finds all providers vulnerable to prompt injection via structural credential and config handling flaws.NeurIPS 2026Must read · vote to see score

91%?TerminalWorld: Benchmarking Agents on Real-World Terminal TasksTerminalWorld automatically builds terminal benchmarks from wild recordings, yielding 1,530 tasks where top agents achieve only 62.5% success with weak correlation to expert benchmarks.NeurIPS 2026Must read · vote to see score · ▲ 8 on HF

91%?Neuro-Inspired Inverse Learning for Planning and ControlInverse Learning trains forward/inverse models and hierarchical stacks for planning and control, matching offline RL and diffusion baselines on D4RL with far less compute while yielding smoother, near-optimal trajectories.NeurIPS 2026Must read · vote to see score

91%?LITMUS: Benchmarking Behavioral Jailbreaks of LLM Agents in Real OS EnvironmentsLITMUS benchmarks LLM agent behavioral jailbreaks in real OS environments, revealing agents execute 40.64% of high-risk operations despite refusals and suffer pervasive execution hallucination.NeurIPS 2026 OralMust read · vote to see score

91%?Coherent Hierarchical Multi-Label Learning to Defer for Medical ImagingCoherent hierarchical multi-label learning to defer uses selective-exclusion contracts to eliminate taxonomic deferral incoherence in medical imaging via projection and belief propagation.NeurIPS 2026Must read · vote to see score

Atlanta Poster Session 5

Fri, Dec 11, 10:00 AM–1:00 PM local time · Hall C1

93%?Neglected Free Lunch from Post-training: Progress Advantage for LLM AgentsRL post-training yields progress advantage, a log-ratio that recovers optimal step-level advantage without dedicated reward models, outperforming trained alternatives across agent benchmarks.NeurIPS 2026Must read · 100% of 1 readers upvoted · vote to see score · ▲ 13 on HF

91%?PACE: Two-Timescale Self-Evolution for Small Language Model AgentsPACE uses two-timescale self-evolution to let frozen small language models improve agents via validated prompt and control updates, outperforming baselines on 12 settings by up to 9.2%.NeurIPS 2026Must read · vote to see score

91%?SALART-VQA: Diagnosing Whether VLMs Understand Salient Artifacts in Generated ImagesSalArt-VQA evaluates VLM artifact understanding via fine-grained questions, revealing high detection recall but only 53% fully correct reasoning and a sensitivity-calibration tradeoff.NeurIPS 2026 OralMust read · vote to see score · ▲ 1 on HF · 4 comments

91%?Learning the Signature of Memorization in Autoregressive Language ModelsFine-tuning produces an invariant memorization signature across architectures that enables transferable learned membership inference achieving over 0.93 AUC on unseen state-space, linear attention, and recurrent models.NeurIPS 2026Must read · vote to see score

Jailbreaks & red teaming89%?Measuring Safety Alignment Effects in Autonomous Security AgentsSafety alignment effects in autonomous security agents require system-level measurement of refusal, tool reliability, and evidence grounding rather than refusal rates alone, with uncensored Gemma models improving security task success but showing mixed, family-dependent effects.NeurIPS 2026Must read · vote to see score

89%?A World Model of Radiologist Reading for Medical Image Representation LearningGazeWorld models radiologist eye-tracking as fixation trajectories through images to pretrain medical representations that achieve state-of-the-art diagnostic and gaze prediction accuracy without requiring real gaze data at inference.NeurIPS 2026Must read · vote to see score

Atlanta Poster Session 6

Fri, Dec 11, 4:30 PM–7:30 PM local time · Hall C1

92%?MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP ServersMCP-Atlas benchmarks LLM tool-use on 1,000 real-server tasks, finding frontier models reach 82.2% pass rates but 63.3% of failures are cognitive.NeurIPS 2026Must read · vote to see score · ▲ 1 on HF

92%?MemLeak: Diagnosing Information Leaks in Multimodal Agent MemoryMultimodal AI agents retain forgotten facts via implicit visual cues, with MemLeak showing 12% image-based recovery and content-aware deletion reducing residuals to 2%.NeurIPS 2026Must read · vote to see score

91%?TraXion: Rethinking Pre-training Frameworks for Mobility and BeyondTraXion introduces MESES axioms and a pre-training framework for multi-entity spatiotemporal event streams that beats mobility baselines and generalizes to security and health logs.NeurIPS 2026Must read · vote to see score

Watermarking & AI-text detection91%?Auditing Cross-Lingual Fairness in Language Model WatermarkingCross-lingual watermark evaluation reveals structural fairness gaps across typological language families rather than isolated language failures.NeurIPS 2026Must read · vote to see score

91%?Reward Bias Substitution: Single-Axis Bias Mitigations Redirect Optimization PressureSingle-axis reward-model bias mitigations redirect optimization onto correlated proxies rather than eliminating it, and auditing on induced distributions with multi-bias tracking is required to certify success.NeurIPS 2026Must read · vote to see score

91%?DRScaffold: Boosting Dense-Scene Reasoning in Lightweight Vision Language ModelsDRScaffold improves lightweight vision-language model reasoning via structured four-stage supervision, surpassing a frozen 32B model on DRBench.NeurIPS 2026Must read · vote to see score