NeurIPS 2026
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 9
- HKUST and Huawei's BalanceRoute lifts LLM decode throughput about 15%
A research team at HKUST and Huawei's 2012 Labs Theory Department jointly proposed BalanceRoute to ease the data-parallel load-balancing bottleneck in LLM serving, where decode workers move in lockstep and the busiest worker forces the others to idle. The method models the synchronization wait mathematically and turns short-term load prediction into online routing decisions; in a 96-card Ascend 910C main experiment running DeepSeek-V3 671B (W8A8), the practical-predictor BR-H improved decode throughput by 15.4% and 14.5% over the strongest baselines on two datasets, while also cutting tail inter-token latency. The paper reports median per-round scheduling latency of 1.20 ms (p99 2.77 ms), says the work was accepted to NeurIPS 2026, and notes it has been deployed in the jointly built MaaS inference system MathBot-Pie.
机器之心 · 🔥 4 - EVA-Cap aligns audiovisual captions to events, lifting ChronusAV by 9%
EVA-Cap is a new framework for audiovisual video captioning that decomposes captions into atomic audiovisual events, replacing unstructured text generation with event-centric alignment. The authors train it with supervised fine-tuning and GRPO on a self-curated dataset, EVA-Data, optimizing three event-centric targets: global event completeness, intra-event attribution and inter-event synchronization. The paper reports a 9.0% average gain on ChronusAV event-centric evaluation and a 3.3% gain on the general WorldSense benchmark. It was accepted at NeurIPS 2026, and the code is said to be released at the link given in the paper.
Hugging Face · Papers · 🔥 0 - OmniVerifier-C1: decomposed graded verifier for rule-governed world modeling
An arXiv paper posted on 15 Oct 2026 introduces OmniVerifier-C1, a visual verifier that moves beyond binary True/False judgments by automatically decomposing complex prompts or environmental rules into fine-grained components and assigning pointwise scores for a continuous, more informative critique signal. The authors train it with a decoupled RL strategy that jointly optimizes sub-question-level binary judgment and global alignment scoring, and report that the pointwise graded verifier significantly outperforms binary verifiers in traditional text-to-image scenarios. They extend the work to rule-governed generative world modeling, arguing theoretically that pairwise verification yields more accurate critiques than the pointwise formulation when ground-truth images are available, with experiments on structured tasks such as Maze and Sudoku; the paper is listed as NeurIPS 2026.
Hugging Face · Papers · 🔥 0 - Apple's LensVLM compresses long context into images, then zooms in
Apple's foundation-model team and Duke University propose LensVLM, which first lets a model skim compressed page images that cover the full text, then calls a tool to expand the pages relevant to the question and read the original content. The paper reports 68.9% average accuracy across seven text QA benchmarks at 4.3x effective compression, close to the 72.4% full-text reference; under 15x initial compression the base model scores 18.1% versus 52.1% for LensVLM. Built on Qwen3.5-9B-Base, it uses Qwen3.5-397B to generate tool-use trajectories for supervised fine-tuning, followed by DAPO reinforcement learning; the work was accepted to NeurIPS 2026 and a LensVLM 9B demo is online.
机器之心 · 🔥 0 - τ-Emotion-Bench exposes LLM agent fragility under emotional users
A paper accepted to the UserSim Workshop at NeurIPS 2026 introduces τ-Emotion-Bench, a ten-domain benchmark that evaluates tool-calling agents under controlled emotion-conditioned user interactions while preserving the underlying task objective. The authors also present Tracer, a scalable agentic data-generation pipeline that builds policy-valid tasks and collects trajectories, letting dissatisfied users terminate inefficient interactions so trajectories with excessive user dependence are filtered out. Experiments with two user simulators show frontier agents remain unreliable, with substantial gaps between average task success and consistent task completion, while Tracer-generated supervision improves Qwen3 models across scales and transfers to τ-Bench and τ-Trait.
Hugging Face · Papers · 🔥 0 - Urbex: agentic spatial grounding for city-scale 3D scenes
A new arXiv paper, "Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes" by Yu Wang, Rui Dai, Wei Chen and five other authors, reframes city-scale 3D grounding from candidate ranking into active spatial search. Urbex represents a city-scale scene as an interactive multi-view environment in which a vision-language model agent can query landmarks, zoom into local regions, render oblique views and commit a grounded bounding box; the tool-use policy is optimized with reinforcement learning on CityRefer using smooth localization rewards plus lightweight shaping rewards for exploration. The authors report the best Hit@1 and IoU on CityRefer, candidate-free grounding without ground-truth boxes, substantially faster performance than the strongest city-scale ranking baseline, and improved cross-scene robustness in zero-shot results on STPLS3D-Refer. The paper is listed as accepted to NeurIPS 2026 (23 pages, 10 figures).
Hugging Face · Papers · 🔥 0 - NeurIPS 2026: attention as an FFN — BUPT's MiTA rethinks efficient attention
Mixture-of-Top-k Attention (MiTA), proposed by a Beijing University of Posts and Telecommunications team, has been accepted at NeurIPS 2026, with the paper (arXiv 2602.01219) and code released. The work expands standard attention into a fast-weights FFN whose weights are generated by the current input and whose hidden width grows with sequence length, and uses that framing to organise prior efficient-attention methods into compression and routing routes plus a five-dimension taxonomy. MiTA derives m landmark queries by average pooling to build m reusable Top-k routed experts plus one shared expert, reaching 79.8% Top-1 on ImageNet-1K with DeiT-S versus 79.9% for full attention; the authors report 58.91% average accuracy on LRA (58.19% for full attention) with total training time cut from 10.6 to 2.4 hours on a single 24GB RTX 4090, and roughly 19x, 56x and 159x inference throughput at 64K, 256K and 1024K sequence lengths.
36氪 人工智能 · 🔥 0 - FramePack-U packs hour-long video into a fixed token budget at constant memory
The paper "Elephant in the Fridge: Constant-Memory Frame Packing for Long Video Understanding" introduces FramePack-U, a training-free frame-packing framework that compresses arbitrarily long videos into a fixed-size visual representation at constant memory cost, requiring no fine-tuning or architectural change and plugging directly into off-the-shelf vision-language models (VLMs) at inference time. It scores each frame by semantic similarity to the text prompt, ranks frames in a diversity-driven manner, and packs the selected frames at varying resolutions into a fixed-length context window so the memory budget goes to frames that are both instruction-relevant and visually diverse. The authors report that on LongVideoBench, MLVU and Video-MME, FramePack-U consistently beats full-resolution baselines and prior token-reduction and keyframe-selection methods, including gains of +12.9%, +10.6% and +9.7% over the official model on Qwen3.5-35B-A3B with a 128-frame budget.
Hugging Face · Papers · 🔥 0 - SkillForge co-evolves agent skills and policy via a fitness-driven skill lifecycle
A NeurIPS 2026-accepted arXiv paper introduces SkillForge, an agentic RL method that evolves a skill library through a fitness-driven lifecycle of trial, active, stable and retired states, so skills and the model co-evolve during training. A pre-RL phase uses the base model's own rollouts to pre-retire low-fitness skills, seeding supervised fine-tuning; RL then continues with selective retirement, stabilization and LLM-guided mutation each iteration. The authors report the highest aggregate success rate across several interactive agent benchmarks, up to 7.8% relative improvement over the strongest baseline while keeping the library compact, and release SkillFurnace, a 5k+ record annotated dataset of retirement-filtered SFT trajectories, evolved skill libraries and retirement events.
Hugging Face · Papers · 🔥 0
Experience and discussion from the community
Share my NeurIPS 2026 experienceAsk about NeurIPS 2026
Nobody has shared their experience with NeurIPS 2026 yet.