Tag

Reinforcing Dual-Path Reasoning in Spatial Vision Language Models

Provides a dual-path approach for spatial vision-language models: a Language-Only Reasoning (LOR) path for stepwise linguistic deduction and a Detect-Then-Reason (DTR) path that detects 3D cues via region tokens before numerical inference. Trains with chain-of-thought cold-start supervision and reinforcement learning to improve 3D grounding and multi-step spatial reasoning.

Beyond the Current Observation: Evaluating Multimodal Large Language Models in Controllable Non-Markov Games

Evaluates multimodal LLMs' ability to reconstruct past observations and act in controllable non-Markov games. Introduces RNG-Bench with two games (Matching Pairs, 3D Maze), three controllable difficulty axes, a head-to-head duel protocol, and a Memory Gap metric to separate forgetting from action errors.

TurboServe: Serving Streaming Video Generation Efficiently and Economically

Serves interactive, long-lived streaming video-generation sessions by jointly scheduling session placement and GPU autoscaling to meet tight per-chunk latency. Combines migration-aware placement, load-driven autoscaling, coalesced chunk processing, GPU–CPU offloading and NCCL GPU–GPU migration; reports ~37% reductions in worst-case per-chunk latency and GPU operating cost.

arXiv LaTeX Source Dataset

Provides pre-parsed arXiv LaTeX source files aligned with official metadata as ready-to-query Parquet rows. Bundles each paper's .tex/.bib/.sty etc. into a single readable tree, updates monthly, and simplifies large-scale access for LLM pretraining, document understanding, and citation analysis while requiring adherence to original arXiv licenses.

Syn4D: A Multiview Synthetic 4D Dataset

Provides multiview synthetic RGB video clips with per-frame depth, instance masks, dense long-range 3D point tracks, camera poses, and SMPL‑X human pose/shape labels for 4D reconstruction, tracking, and geometry-aware novel-view synthesis. Includes ~4.7K clips (1.4M frames) and is licensed for AI training.

PerceptionRubrics: Calibrating Multimodal Evaluation to Human Perception

Provides a rubric-based benchmark that converts dense image captions into instance-specific atomic checks (Must-Right and Easy-Wrong) and a gated scoring rule, aiming to expose perceptual brittleness and better align multimodal model evaluation with human judgment.

The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning

Proposes Monotonic Inference Policy Improvement (MIPI) and a two-step Monotonic Inference Policy Update (MIPU) to address training–inference probability mismatch in LLM reinforcement learning by constructing sampler-referenced candidate updates and accepting synchronized updates using an inference-gap proxy; shows improved reasoning accuracy and stability under FP8-quantized rollouts.

Morphing into Hybrid Attention Models

Treats hybrid layer selection as a budget-constrained subset optimization and introduces FlashMorph: a pipeline that equips each transformer layer with a linear-attention branch, jointly optimizes layerwise gates on synthetic long-context retrieval data, then discretizes, distills, and finetunes—achieving strong long-context recall using only 20M selection tokens.

AGE: Adaptive-masking for Graph Embedding in Graph Retrieval-Augmented Generation

Trains a transformer-based graph encoder with RL-guided adaptive masking so retrieved subgraphs embed relationships that better align with frozen LLM text encoders, improving GraphRAG performance with non-parametric retrievers on GraphQA benchmarks.

MemSyco-Bench: Benchmarking Sycophancy in Agent Memory

Evaluates how long-term memory in LLM agents amplifies sycophantic behavior and when memory should or should not influence decisions. Provides five targeted tasks, 1,550 standardized samples, an evaluation pipeline, and baseline adapters to test memory use, conflicts, scope, updates, and personalization.

Domain Arithmetic: One-Shot VLA Adaptation under Environmental Shifts

Adapts pretrained Vision-Language-Action (VLA) models to new camera poses and robot embodiments from a single demonstration by performing weight-vector arithmetic that injects domain-specific information. Filters noise via subspace alignment of singular components; designed for one-shot adaptation under visual and embodiment shifts.

ELDR: Expert-Locality-Aware Decode Routing for PD-Disaggregated MoE Serving

Predicts per-request MoE expert footprints from prefill activations and routes decode requests to workers that maximize expert-locality, lowering decode latency by combining offline K-means partitioning with online locality-band routing and a KV-block–coindexed signature cache.