Daily AI Papers — November 08, 2026

1. Qwen3-TTS Technical Report

Authors: Hangrui Hu, Xinfa Zhu, Ting He, Dake Guo, Bin Zhang, Xiong Wang, Zhifang Guo, Ziyue Jiang, Hongkun Hao, Zishan Guo, Xinyu Zhang, Pei Zhang, Baosong Yang, Jin Xu, Jingren Zhou, Junyang Lin arXiv: arxiv.org/abs/2601.15621 Summary: Qwen3-TTS is a multilingual streaming text-to-speech family trained on more than five million hours of speech across ten languages, with three-second voice cloning and description-based voice control. The report describes two speech tokenizers and a dual-track language-model architecture, including a low-bitrate configuration with 97-millisecond first-packet emission, and releases the models and tokenizers under Apache 2.0.

2. Tools Need Not Speak in Text: Latent Context Compression for Agents

Authors: Yuetong Liu, Zhehao Zhou, Yitong Zhang, Haorui Chen, Jia Li arXiv: arxiv.org/abs/2611.06500 Summary: Coca-Cola compresses tool observations into continuous latent tokens and uses Action-Preserving Training to preserve an agent’s next-action distribution rather than retaining raw text. Across four question-answering benchmarks and two model families, the authors report 70.01%-89.24% lower token consumption and 24.45%-55.97% lower end-to-end runtime while improving task performance.

3. Signed but Steerable: Measuring and Closing the Text-Vector Integrity Gap in Offline Retrieval-Augmented Generation

Authors: Georgios Feretzakis, Alexandros Papaspyridis arXiv: arxiv.org/abs/2611.06498 Summary: This study shows that signatures on offline RAG knowledge packs do not guarantee that their embedding vectors faithfully encode their text, and that cosine-tolerance checks can leave retrieval vulnerable to tampered vectors. The authors demonstrate attacks on edge-device and compressed-vector-store settings, then evaluate verifier-controlled quantization codes, guard bands, and canonicalization to remove the publisher’s choice within the tested integrity checks.

4. Identically Trained Language Models Break on Different Jailbreaks

Authors: Chi-Wei Chang, Mu-Tien Kuo, Kuan-Wei Lee, Meng-Hsun Tsai arXiv: arxiv.org/abs/2610.34074 Summary: Across 150 identically safety-post-trained models spanning 25 data recipes, this study finds that the particular jailbreaks a checkpoint fails depend strongly on its pretraining seed. Per-prompt non-refusal correlations fall from 0.984 for copies of one pretraining run to 0.811 across pretraining seeds, suggesting that one-checkpoint jailbreak evaluations miss important training-run variability.

5. NPCBench: A Clinical Apprenticeship Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma

Authors: Pengkai Wang, Wei-Wei Zhang, Yan Li, Min Tang, Zhitian Hou, Zeyu Liu, Guanghao Zhu, Yuanyi Wang, Yanggan Gu, Wenjun Wang, Minheng Ni, Congkai Xie, Zhijie Sang, Jianmin Wu, Ying Sun, Hongxia Yang arXiv: arxiv.org/abs/2611.08369 Summary: NPCBench evaluates nasopharyngeal-carcinoma care as a clinical apprenticeship, combining 4,777 component and consultation items with a 100-patient longitudinal CarePath cohort. Across 30 models, strong isolated-task scores do not ensure jointly correct diagnosis and care, with strict CarePath success reaching at most 23% among the 21 multimodal models.

6. UniRestorer++: Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling

Authors: Jingbo Lin, Zhilu Zhang, Fan Li, Renjing Pei, Wenbo Li, Hang Xu, Hongzhi Zhang, Wangmeng Zuo arXiv: arxiv.org/abs/2611.06042 Summary: UniRestorer++ scales universal image restoration by refining degradation groups into specialized experts and routing at both path and granularity levels, with spatially dense representations for nonuniform corruption. On the new CDD-50 benchmark and other evaluations, the authors report approximately log-linear PSNR scaling, stronger out-of-distribution gains than vanilla MoE scaling, and zero-shot and few-shot transfer to real-world images.

7. Hard Sample-Conditioned Decorrelated Attention for Robust Face Recognition

Authors: Tahar Chettaoui, Guray Ozgur, Eduarda Caldeira, Naser Damer, Fadi Boutros arXiv: arxiv.org/abs/2610.41789 Summary: ViT-HarDis strengthens attention-head decorrelation on difficult face-recognition samples whose embeddings resemble competing identities, encouraging specialization where ambiguity is greatest. The authors report improvements over baseline and uniform regularization on most benchmarks, including leading results among ViT-based methods on IJB-B, IJB-C, and TinyFace without additional inference cost.

8. HERIT: Enabling Global Access to Hanja-to-English Historical Translation by Mitigating Temporal Bias

Authors: Dayeon Jeong, Kyeongpil Kang arXiv: arxiv.org/abs/2611.04949 Summary: HERIT uses retrieval-augmented generation to create Hanja-to-English pseudo-labels from abundant Hanja-Korean archives, expanding domain-adaptation data across historical periods. Automated and human evaluations report gains over baselines, and the resulting model translates previously untranslated archive material to improve global research access.

9. SynthFaces: Balanced Large Scale Human Dataset

Authors: Sergey Karpukhin, Denis Bobkov, Kirill Morozov, Aibek Alanov arXiv: arxiv.org/abs/2610.40119 Summary: SynthFaces combines an Arc2Face diffusion model with flow matching over identity embeddings to generate a balanced, identity-centric synthetic human dataset. The authors evaluate identity separation with a nearest-neighbor non-leakage metric and demonstrate the dataset’s usefulness in fine-tuning an adapter-based personalized-generation method.

10. DRAP: Dynamics-Risk Adaptive Plasticity for Continual Learning in Physical AI

Authors: Oswaldo Ludwig arXiv: arxiv.org/abs/2611.05029 Summary: DRAP is a replay-free continual-learning method for physical AI that weights historical action-gradient importance by a local estimate of trajectory-error amplification and adjusts each tensor’s AdamW learning rate accordingly. Sequential OpenVLA adaptation across LIBERO Spatial, Object, and Goal suites reports better retention of prior behaviors than standard fine-tuning and EWC while preserving strong acquisition of new tasks.

11. Scientist World: Grow a Research World for AI-Scientists

Authors: Taiming Lu, Sachin Konan, Zhuang Liu arXiv: arxiv.org/abs/2611.08478 Summary: Scientist World turns research papers into executable environments by extracting open-ended questions and reproducing each paper’s method as a baseline, releasing an initial 1,000 verified environments. The authors report frontier AI-scientists surpassing the human baselines in 74% of evaluations, collect 6,400 hours of research traces, and use the environments to improve a smaller model’s research ideation.

12. Benchmarks as Software: Terminal-Bench and the Case for Continuous Benchmarks

Authors: E. Kelly Buchanan, Crystal Zhou, Scott W. Linderman, Christopher Ré arXiv: arxiv.org/abs/2611.08417 Summary: This paper introduces a continuous-validation workflow for agent benchmarks and applies it to Terminal-Bench 2.0, finding issues in roughly 30% of tasks and releasing repairs as Terminal-Bench 2.1. Repairs increase measured accuracy by more than ten percentage points and shift leaderboard positions, while time-budget experiments show that resource allocation must also be validated as part of benchmark maintenance.

13. StrataFlux: Memory-Hierarchy-Native Low-Bit LLM Inference

Authors: Jake Hyun, Ziyu Zhao, Jae-sun Seo, Mohamed S. Abdelfattah arXiv: arxiv.org/abs/2611.08363 Summary: StrataFlux selectively streams additional weight and KV-cache precision from host memory or storage, overlapping transfers with computation to improve low-bit inference without making every decoding step wait. On the tested hardware and models, it improves accuracy while retaining most quantized throughput, including 47.1% versus 11.0% AIME accuracy for Qwen3-8B with two-bit weights and KV at unchanged GPU memory use.

14. SpecCTMC: Exact Speculative Simulation of Diffusion Language Model Continuous-Time Markov Chains

Authors: Siva Rajesh Kasa, Yasong Dai, Sumit Negi, Hongdong Li arXiv: arxiv.org/abs/2611.08362 Summary: SpecCTMC speculates successive token updates from cached model predictions, checks intermediate states in parallel, and applies exact acceptance and correction for diffusion-language-model sampling. The authors prove preservation of the underlying continuous-time Markov chain’s joint sampling distribution and report 1.73-4.38-times fewer sequential model evaluations across five benchmarks without a separate draft model.

15. StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning

Authors: Yang Ouyang, Jung-Eun Kim arXiv: arxiv.org/abs/2611.08403 Summary: StepBack identifies likely first errors in a small set of draft reasoning trajectories and resamples the error step and its suffix instead of repeatedly generating complete solutions. Across seven mathematical-reasoning and multi-hop question-answering benchmarks, it matches self-consistency accuracy on average while reducing token use by up to 41.26% and can be combined with adaptive inference methods.

16. Beyond K/V Caches: Cross-Model State Handoff in Hybrid Linear-Attention LLMs

Authors: Jiacheng Cai, Tao Zhang, Siqi Liang, Yun Dai, Ziqian Zeng arXiv: arxiv.org/abs/2611.08133 Summary: Hybrid State Transfer maps full-attention caches, recurrent states, and convolution histories between hybrid linear-attention models, calibrating adapters against the frozen target’s continuation outputs. Across five benchmarks, Qwen3.5 handoffs from 4B and 9B to 27B retain 89.88% and 94.11% of mean task performance, with separate synthetic-16K timing showing roughly 21.7-times handoff-body speedups when source caches are already available.

17. EgoGeneralist: Unifying Egocentric Action-Video Tasks through Video-Native Modeling

Authors: Yanlin Jin, Bowen Yang, Cheng Zhang, Yifan Yang, Zhe Li, Jianfei Cai, Kai Zhu arXiv: arxiv.org/abs/2611.08477 Summary: EgoGeneralist adapts pretrained video generators with a role-adaptive action renderer and cross-stream coupling to unify five egocentric action-video tasks, including generation and pose estimation. Despite adaptation only on controlled laboratory recordings, the authors report generalization to unseen scenes and interactions while linking metric three-dimensional motion to strong pretrained visual priors.

18. PlayGen2: Learning Real-Time Playable Worlds with Explicit 3D State

Authors: Jihai Zhang, Yuhan Zhou, Junyou Li, Deheng Ye, Qi Dou, Yu Cheng arXiv: arxiv.org/abs/2611.07867 Summary: PlayGen2 separates persistent three-dimensional world-state evolution from neural rendering, maintaining geometry and identity-indexed entities rather than feeding generated frames back into dynamics prediction. In Doom experiments, it supports unseen maps, composed multiplayer interaction, and programmable state interventions, with the complete simulation-and-rendering pipeline reaching approximately 160-millisecond backend latency on an M4 Pro MacBook Pro.

19. Agent-R$^2$: Scaling Agents for Realistic Productivity Work via Real-Source Data Synthesis

Authors: Teng Pan, Zixuan Wang, Yuchen Yan, Qifei Wu, Yao Chen, Yingqi Qu, Ruiqing Zhang, Yongliang Shen arXiv: arxiv.org/abs/2611.08219 Summary: Agent-R² synthesizes realistic productivity training tasks from occupational seeds, corpus-derived workflows, and authentic web files, distilling 4,000 instances with deliverable-level rubrics. Supervised fine-tuning and rubric-based reinforcement learning produce a 4B agent that exceeds the compared Qwen3-235B model on GDPval and OfficeQA while reducing premature, insufficiently grounded deliverables.

20. Spectral and Subspace Dynamics of LoRA in Supervised Fine-Tuning and Reinforcement Learning

Authors: Chenruo Liu, Zihan Wang, Yijun Dong, Qi Lei arXiv: arxiv.org/abs/2611.08414 Summary: This study tracks LoRA update spectra and subspaces during supervised fine-tuning and reinforcement learning, finding that group-mean baseline subtraction explains RL’s stronger spectral concentration and faster subspace rotation. Stage-specific update transformations encourage broader adaptation in SFT and rotation in RL, yielding the best performance among compared methods in most tested mathematical-reasoning settings.