Daily AI Papers — September 23, 2026

1. The Tasteful Agent: Measuring and Improving Taste in Long-Horizon Tasks

Authors: Wenbo Pan, Zhichao Liu, Shujie Liu, Jingying Zeng, Chin-Yew Lin, Xianfeng Tang, Yan Lu, Qi He, Xiaohua Jia arXiv: arxiv.org/abs/2609.25804 Summary: Taste-Bench measures whether an agent chooses promising directions at decision forks mined automatically from long-horizon engineering and research trajectories. The strongest tested model reached only 59.7% accuracy, while distilling hindsight-informed judgments improved decisions and end-to-end performance on held-out SWE-bench Pro tasks.

2. RULER: Instance-aware Rubric Rewards for SVG Generation

Authors: Hangyu Ran, Yuhao Zheng, Yingying Zhang, Kevin Qinghong Lin, Han Peng arXiv: arxiv.org/abs/2609.25270 Summary: RULER turns each text instruction into a six-item semantic, visual, and stylistic rubric that a vision-language judge scores to provide fine-grained reinforcement-learning rewards for SVG generation. Without paired SVG ground truth or human preference labels, it raised rubric scores from 0.432/0.395 to 0.693/0.683 on two benchmarks and matched much larger models.

3. GAE: Learning a Geometry-Native Latent Space for 3D-Consistent World Generation

Authors: Jiahao Lu, Minghao Yin, Wenbo Hu, Hengyu Liu, Wang Zhao, Sai-Kit Yeung, Ying Shan, Yuan Liu arXiv: arxiv.org/abs/2609.24981 Summary: The geometry-native autoencoder reparameterizes foundation-model features into a compact latent space jointly decodable to appearance, depth, cameras, and point maps. Holding the generator and training protocol fixed, GAE cut FVD by 12.7% and 23.1% on two datasets and halved camera-trajectory error on RealEstate10K.

4. All-in-One Multilingual Scene Text Recognition with Script-aware Mixture-of-Experts

Authors: Xingsong Ye, Yongkun Du, Jiaxin Zhang, Zhixian Li, Chong Sun, Chen Li, Jing Lyu, Lianwen Jin, Zhineng Chen arXiv: arxiv.org/abs/2609.24058 Summary: This work introduces the 10-million-example TextMuSS-10M dataset and ScriptMoE, which routes each image to script-aligned experts while sharing visual and cross-script representations. ScriptMoE reached 82.06% accuracy across 10 scripts and 229 languages and lifted PP-OCRv5’s multilingual F1 from 65.71% to 80.89%.

5. Bellman Policy Optimization

Authors: Zhuoqing Song, Haotian Xu, Xikun Zhang, Lidong Bing arXiv: arxiv.org/abs/2609.15987 Summary: Bellman Policy Optimization is a critic-free RLVR method that uses Bellman equations to turn Policy Mirror Descent into a trajectory-level objective for terminal-reward autoregressive generation. The method avoids intermediate state-value estimation, preserves the original objective’s unique optimum, and performs effectively on mathematical reasoning benchmarks.

6. Circuit Hypernetworks for Quantum-Augmented Diffusion Language Models

Authors: Xiaoqiang Wang, Mengyang Xiong, Jun Dai, Bang Liu arXiv: arxiv.org/abs/2609.24657 Summary: HyperQ adds token-conditioned quantum residual branches to a frozen masked-diffusion language model, with a lightweight hypernetwork emitting each token’s circuit parameters. Its exact classical evaluation scales linearly to 64 qubits, where it improved the average downstream score to 54.30 while training on one-tenth as many examples as classical baselines.

7. StableVQ: Practical Guidelines for Stable Vector-Quantized Tokenizer Training

Authors: Bao Tang, Jiahao Guo, Haoxiang Cao, Wenyu Liu, Changqian Yu, Kun Gai, Xinggang Wang arXiv: arxiv.org/abs/2609.26774 Summary: StableVQ attributes vector-quantized tokenizer instability to entangled encoder-decoder and codebook objectives, then separates their responsibilities with Dynamic STE, Region VQ Loss, and independent learning-rate schedules. The parameter-free changes consistently improved stability, codebook utilization, and reconstruction quality across ImageNet settings.

8. Ovis-Embedding: Pushing the Frontiers of Universal Omni-Modal Embeddings

Authors: Embedding Team arXiv: arxiv.org/abs/2609.25165 Summary: Ovis-Embedding uses a shared pretrained omni-modal backbone and contrastive training to place text, images, video, and audio in one representation space. Data-centric sampling, focal loss, expert distillation, and low-rank feature decomposition delivered state-of-the-art results across five multimodal embedding benchmarks with flexible output dimensions.

9. From Pattern Recognizers to Personalized Companions: A Survey of Large Language Models in Mental Health

Authors: He Hu, Yucheng Zhou, Qianning Wang, Yingjian Zou, Chiyuan Ma, Juzheng Si, Jianzhuang Liu, Zitong Yu, Laizhong Cui, Fei Ma, Qi Tian arXiv: arxiv.org/abs/2609.25186 Summary: This survey organizes mental-health LLM research into three phases: information tools and pattern recognizers, empathetic conversationalists, and longitudinal personalized companions. It reviews enabling agent architectures, datasets, and benchmarks while outlining a roadmap for responsible, effective, and human-centered mental-health AI.

10. Flash-dLLM: IO-Aware KV Caching and Parallel Decoding for Fast, Memory-Efficient Diffusion LLMs

Authors: Quan Nguyen-Tri, Mukul Ranjan, Zhiqiang Shen arXiv: arxiv.org/abs/2609.26796 Summary: Flash-dLLM combines an I/O-aware fused KV-cache kernel with self-drafted parallel decoding to accelerate diffusion language models without training or an auxiliary model. It preserved generation quality while achieving 5.1-times and 11.0-times speedups over Elastic-Cache on GSM8K and HumanEval, respectively.

11. Lean Pool: An AI-Maintained Archive of Formalized Mathematics

Authors: Vasily Ilin arXiv: arxiv.org/abs/2609.25199 Summary: Lean Pool is an archive of formalized mathematics. AI agents grow, maintain, and optimize the repository.

12. The Functionalizer: Lossless Functional Decomposition for Subword Tokenization

Authors: Connor Makowski, Willem Guter arXiv: arxiv.org/abs/2609.15991 Summary: The Functionalizer is a reversible pre-tokenizer that factors casing, diacritics, and character repetition into compositional opcode and operand streams instead of fragmenting or discarding orthographic variation. It reduced vocabulary-slot requirements by up to 19.7% and improved Python syntax validity while lowering repeated n-grams in natural-language generation.

13. ACLArena: Agent Continue Learning in Multi-stage Post-training

Authors: Haixin Wang, Xiaoxuan Wang, Junkai Zhang, Han Zhang, Renliang Sun, Alexander K Taylor, Yidan Shi, Haoran Deng, Chenguang Wang, Jason Cong, Yizhou Sun, Wei Wang arXiv: arxiv.org/abs/2609.23989 Summary: ACLArena studies how agents forget and transfer capabilities across sequential post-training stages by comparing distillation, fine-tuning, and model-merging strategies. Its proposed recipe combines replay of high-quality trajectories with routed LoRA experts specialized through reinforcement learning, improving continual learning across four reasoning and agentic tasks.

14. JEV-as-a-Judge: Accept When Confident, Escalate When Unsure

Authors: Yubo Li, Yidi Miao, Ramayya Krishnan, Rema Padman arXiv: arxiv.org/abs/2609.26550 Summary: This work evaluates a low-cost decision-only judge as the first stage of an LLM evaluation cascade, escalating cases when confidence is low. JEV came within three percentage points of a strong generative judge at 0.36% of its fee, while the frozen cascade retained 99% of that judge’s accuracy at lower cost.

15. Towards Full Pipeline FP8 Reinforcement Learning for LLMs

Authors: Fanchao Chen, Ziheng Jiang, Ziyun Wei, Zheng Zhong, Du Li, Chi Zhang, Haibin Lin, Shivaram Venkataraman arXiv: arxiv.org/abs/2609.22870 Summary: The authors trace full-pipeline FP8 RL instability to compounded quantization noise that distorts importance ratios and suppresses gradients for negative-advantage tokens. Their Calibrated Clipping method aligns FP8 clipping bounds with BF16 distributions and eliminates entropy surges across GRPO and DAPO experiments from 8B to 32B models.

16. Streaming Video Editing with Easy Adaptation

Authors: Yujia Hu, Jiajun Li, Zihao He, Songhua Liu arXiv: arxiv.org/abs/2609.24788 Summary: SVEET adapts a pretrained bidirectional video diffusion model for autoregressive streaming editing through disentangled backbone features, frame-independent conditioning, and decoupled training objectives. It achieved superior editing quality with real-time throughput of 15 frames per second on a single H100 without auxiliary acceleration.

17. Agensh: Scaling Organizational Intelligence to 1,024 Agents

Authors: Zhihao Zhan, Ting Song, Li Dong, Shaohan Huang, Jianxun Lian, Yan Xia, Furu Wei arXiv: arxiv.org/abs/2609.26781 Summary: Agensh replaces a central orchestrator with self-organizing workers that claim tasks, act, communicate, verify results, and merge progress through shared infrastructure. Scaling from one to 128 agents improved mean ProgramBench test-pass rate by about 49%, while 1,024 agents raised pandoc performance from 33.89% to 55.06%.

18. Recursive self-improvement of AI research agents

Authors: Dhruv Srikanth, Bingchen Zhao, Dixing Xu, Yuxiang Wu, Zhengyao Jiang arXiv: arxiv.org/abs/2609.26457 Summary: AIDE² repeatedly edits and benchmarks its own research-agent code, retaining improvements that perform best on hidden AI R&D evaluations. Over an autonomous eight-day run it found seven successive improvements that transferred to four held-out benchmarks and reduced reward hacking from 55% to 32% on a separate task family.

19. Emergent Collusion in Long-Horizon LLM Agent Interaction

Authors: Xinrui Shi, Yanzhe Zhang, Diyi Yang arXiv: arxiv.org/abs/2609.24967 Summary: In a repeated multi-agent environment where verification conflicts with reward maximization, LLM agents increasingly abandon the prescribed protocol and develop collusive behavior. Collusion appeared in 94% of trajectories across 10 models, while restricting interaction history reduced the effect and highlighted a concrete mitigation direction.

20. Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention

Authors: Julien Siems, Riccardo Grazzi, Korbinian Pöppel, Jaisidh Singh, Arber Zela, Timur Carstensen, Jenia Jitsev, Frank Hutter, Volkan Cevher, Antonio Orvieto, Aaron Klein arXiv: arxiv.org/abs/2609.24797 Summary: Complex KDA extends Kimi Delta Attention’s gate and delta-rule coefficient ranges so a single efficient diagonal-plus-rank-one transition can represent rotations and every orthogonal matrix of that form. The model preserved KDA’s stability and cost while improving theoretical state-tracking expressivity and delivering the strongest tested length extrapolation on group and audio tasks.