neurips-2026

[NeurIPS 2026] VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon

StepShield: When, Not Whether to Intervene on Rogue Agents — NeurIPS 2026 benchmark for temporal evaluation of AI agent guardrails (9,429 trajectories, step-level annotations)

[NeurIPS 2026] StableHand: Quality-Aware Flow Matching for World-Space Dual-Hand Motion Estimation from Egocentric Video

[NeurIPS 2026] FlowHOI: Flow-based Semantics-Grounded Generation of Hand-Object Interactions for Dexterous Robot Manipulation

[NeurIPS 2026] ReasoningShield: Safety Detection over Reasoning Traces of Large Reasoning Models

NeurIPS 2026: Mechanistic Interpretability toolkit for Vision-Language-Action models

[NeurIPS 2026] Croissant Baker: generate Croissant metadata for ML datasets, straight from a directory of files

[NeurIPS 2026] "CausalCompass: Evaluating the Robustness of Time-Series Causal Discovery in Misspecified Scenarios"

[NeurIPS 2026] FedAgent: A Library for Decentralized Agent Learning. 🏆 Best Paper Award at the AAAI'26 TrustAgent Workshop, 🏆 Outstanding Paper Award in the AAAI'26 PerFM Workshop.

K-Veritas CLI

Code, models and data for M-plicits: Neural Implicit Surfaces via Nested Multiscale Residuals (NeurIPS 2026). A noise-robust neural surface representation built on nested neighborhood bands, with real-time multiscale sphere tracing and analytical normals.

The Neural Interface Foundation Challenge — NeurIPS 2026 Sydney. An open-source brain-decoding benchmark with four tracks across EEG, EMG, sleep, and BCI.

SecureClaw: Clawing Back Control of LLM Agents — NeurIPS 2026. Runtime, method adapters, and evaluation code.

SplitZip: Ultra Fast Lossless KV Compression for Disaggregated LLM Serving

[NeurIPS 2026] KaliBench: A Fine-Grained Benchmark for Cybersecurity Tool Use on Kali Linux with Runtime-Free Verifiable Rewards

interactive, sub-16ms vector explainer built for the DataForge 2026 Pathway Track.It proves why classical State Space Models suffer exponential associative forgetting ($\rho(A)^L$) and visualizes how Pathway's Dragon Hatchling (BDH) synaptic plasticity and BDH CQ continuous latent reasoning eliminate interference without KV-cache explosion.

Official code for "Zero-Shot Quantization via Weight-Space Arithmetic" (NeurIPS 2026). Transfer quantization robustness across tasks with quantization vectors, no receiver-side QAT.

[NeurIPS 2026] MultiBBQ: A Fairness Benchmark for Multimodal LLMs. 🏆 Best Paper Award in the ACL'26 TrustNLP Workshop.

PersonaManifold (NeurIPS 2026): curved geometry, geodesic steering, and behavioral triplet evaluation

Official code for "Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection" (NeurIPS 2026). RAMSAD selects the best anomaly detector for a new series by retrieving similar series from a knowledge base, with no selector training.