AI Research

Researchers posted arXiv:2609.19996v1, a deep architecture that jointly optimizes cost functions and a route-ranking model for customizable route planning. The system runs a multi-objective Dijkstra algorithm offline to …

A new arXiv paper (2609.19934v1) proposes the Depth Control Protocol (DCP), a diagnostic suite that disentangles three factors conflated by depth truncation when evaluating whether depth-recurrent language models genuine…

A study of 67,200 responses from GPT, Claude and Gemini found that the language of a question changes how the models assess the war in Ukraine, with the balance of Russia-leaning and Ukraine-leaning answers varying acros…

Researchers behind the arXiv paper DART: Distillation-Aware Reparameterization introduced a training-free method that combines low-rank coordinate transport with target-schedule response calibration to reuse LoRA adapter…

A cooperative framework called UnifiedPlayers, comprising a Planning Player, an Execution Player, and an Evaluation Player, outperformed the strongest prior baseline by at least 3.5% on mathematical reasoning and 3.9% on…

A new arXiv paper (2609.20068v1) unifies marginal utility, matrix factorization, and the transformer Key-Value cache under a single allocation rule that retains the top dimensions whose eigenvalue exceeds the shadow pric…

A new arXiv paper (2609.20152v1) introduces MTVA-Bench, a benchmark that evaluates the language model inside cascaded voice agents under the same conditions it faces in production, covering 49 agents, 490 reviewed scenar…

A Formal Concept Analysis-guided counterfactual (FCA-CF) framework achieved Validity = 1.0000, Sparsity = 2.37 features changed, and Proximity = 0.900 on 60 benign-predicted TCGA-BRCA instances, according to an arXiv pap…

Researchers introduced JointMatch, a learning-based framework that solves ride-sharing request pairing and vehicle assignment together on a single sparsified graph using a graph neural network. On New York City Yellow Ta…

MAGMA-GEN, an on-policy data-generation pipeline from researchers publishing as arXiv:2609.20056v1, converts ambiguous failed rollouts in hierarchical robotic systems into validated recovery supervision by retaining a pr…

A decomposed confidence layer combining perception, layout, and validation channels with conformal risk control raised AUROC for financial-document key-value extraction from 0.54-0.74 with native Vision Language Model ve…

Nunchux AI researchers Muyang Li and collaborators published the VC-Attention paper on September 14, reporting attention-kernel speedups of 1.46x to 1.59x on NVIDIA H200 and B200 data-center GPUs and up to 1.70x faster e…

VIDRAFT AI Research published a paper (arXiv:2609.20269) testing whether Transformer attention mechanisms need to repeat identically at every layer, using a 7×7 Latin square to spread seven sequence mixers evenly across …

A proposed framework called Natural General Intelligence calls for a foundation model grounded in the state and dynamics of the planet itself, connecting the biosphere, atmosphere, oceans, land, ice, and subsurface to an…

OpenAI's InstructGPT/RLHF work showed that GPT-2-sized models (over 100x smaller than GPT-3) trained on the right task beat GPT-3, according to an essay by the author of The Bitterest Lesson. The essay argues that the pr…

PrismML released Bonsai 2 27B on September 17, 2026, a ternary-quantized version of Alibaba's open Qwen3.8-27B that retains 98.2% of the full-precision model's benchmark average while shrinking from 54GB to 5.9GB. The mo…

Researchers introduced When2Think, a difficulty-aware length control method for hybrid reasoning models that aims to stop large reasoning models from overthinking easy problems and underthinking hard ones. The work targe…

A new method called RetireOPD proposes self-retiring on-policy distillation to give multi-turn reinforcement learning agents dense token-level supervision from a self-teacher with privileged task skills, so a skill-free …

Video DeltaNet introduces a video-native hybrid attention mechanism for livestream video generation, targeting the computational bottleneck caused by repeated processing of long spatiotemporal token sequences during deno…

DeepSeek released DeepSeek-V4.1-Flash, a model that targets KV cache compression to reduce the memory and storage burden of long-context, input-heavy agent workloads. DeepSeek said prefill remains computationally expensi…