HiVe is a prompt‑tuning framework that builds a hierarchy of prompts by exploiting inter‑task relationships during training. It uses a vertical mixture‑of‑experts (V‑MoE) at inference to compose prompts at the level of specialization needed for each input, allowing input‑dependent prompt adaptation. Experiments demonstrate that HiVe consistently outperforms strong prompt‑tuning baselines across diverse tasks.
By HyeonJik Bae, Minyeol Kim, Susik Yoon
arXiv:2608.28611v1 Announce Type: cross
Abstract: Recent advances in large language models (LLMs) like ChatGPT and LLaMA have transformed AI-driven education, but these systems are predominantly trai...
By Isha Narang, Sneh Gosai, Mayank Singh
The paper introduces Juris Policy Optimization (JPO), a post‑training framework designed to enhance structured legal reasoning in Chinese criminal judgment prediction. JPO first trains models with teacher‑generated rationales to guide a four‑step reasoning process, then applies reinforcement learning using a composite reward that balances prediction accuracy, reasoning completeness, and cross‑step consistency. Experiments on several open‑source language models and three Chinese legal benchmarks demonstrate that JPO consistently outperforms both supervised fine‑tuning and standard reinforcement learning baselines in terms of judgment prediction and reasoning quality.
By Zhaolu Kang, Yantao Liu, Tailong Luo, Leqi Zheng, Lei Wei, Chenghua Zhu, Junhao Gong, Jiachen Qian, Eric Hanchen Jiang, Jiaxin Liu, Yuan Wang, Hao Zhang, Zixia Wang, Rong Fu, Zheng Lin, Richeng Xuan, Zhichao Hu
Randomized YaRN is a training method that enhances length generalization for large language models by combining YaRN-based positional extrapolation with randomized positional encoding and a length curriculum. During training on short-context data, tokens receive YaRN positional encodings sampled from a larger position range, exposing the model to out-of-distribution positional representations. Evaluated on BABILong, Multi-Round Coreference Resolution, and LongBench v2, Randomized YaRN consistently improves reasoning performance on context lengths from 16K to 128K, outperforming standard fine‑tuning especially at far out‑of‑distribution lengths.
By Manas Mehta, Fangcong Yin, Greg Durrett
arXiv:2608.29461v1 Announce Type: new
Abstract: Sandbagging models strategically underperform on evaluations while retaining the capabilities being measured. The evaluations that guide frontier-model...
By Hong Kiat Tan, Linh Le, David Williams-King
The paper explores whether structured linguistic reasoning traces can improve low‑resource machine translation by guiding large language models (LLMs). It proposes a pipeline that automatically generates step‑by‑step reasoning traces from Universal Dependencies treebanks, dictionaries, and grammar‑rule banks, and evaluates these traces in in‑context learning, supervised fine‑tuning, and reinforcement fine‑tuning on Xibe and Chintang. The results show that providing reliable reasoning traces at inference time significantly boosts translation quality, whereas using them as training data yields smaller, less consistent gains, indicating that LLMs can benefit from grammatical guidance but struggle to generate accurate analyses themselves.
By Renhao Pei, Yihong Liu, Sampo Pyysalo, Hinrich Sch\"utze, Shaoxiong Ji
The paper investigates how to choose language models (teachers) for generating multilingual synthetic data used to fine‑tune smaller student models. By evaluating 10 teacher models across six diverse languages and training 240 students, the authors find that teacher effectiveness is not driven by model size but by data qualities such as prompt diversity, length, and fluency, which explain most of the variance in student performance. Practical guidelines are offered, including matching teacher and student families and using translated prompts to improve outcomes for low‑resource languages.
By Lester James V. Miranda, Ivan Vuli\'c, Anna Korhonen
arXiv:2608.28809v1 Announce Type: new
Abstract: Extreme low-bit inference offers a route toward smaller models and constrained deployment. Ternary language models restrict weights to $\{-1,0,+1\}$, a...
By Anirudh Malik, M Sparsh Mehra, Poojith Devan
CogEvol is a family of models designed to generate learning environments—structured JSON slides or interactive HTML pages—from course briefs in a single pass. In production, it completes a slide in a median of 17 seconds and an interactive page in 59 seconds, vastly faster than multi‑turn agent scaffolding. The system enforces reliability through a production‑grounded data pipeline that produces over 53,000 verified SFT samples and a hybrid rule‑plus‑VLM reward that guides GRPO‑based reinforcement learning, with a 27B model achieving high quality scores while using far fewer parameters than flagship coding models.
By Shangqing Tu, Daniel Zhang-Li, Yucheng Wang, Shiyu Gan, Yanpeng Wang, Huiqiang Rong, Mofei Chen, Shen Yang, Yini Chen, Yinuo Duan, Haoxuan Li, Binglin Liu, Ye He, Danqi Zheng, Zhanxin Hao, Yuxuan Wu, Mengting Tao, Yuqiu Liu, Jifan Yu, Juanzi Li, Bin Xu, Lei Hou, Huiqin Liu, Yu Zhang
The paper introduces a German-English benchmark dataset to evaluate anti‑LGBTQ biases in language models, combining community‑sourced stereotypes from German‑speaking queer individuals with a German translation of WinoQueer. Eight language models of varying sizes and architectures were assessed, revealing that they reproduce anti‑queer stereotypes with differences across identities and models. Fine‑tuning on community and progressive media content reduced bias on average, though the effect was not consistent across all models and identities.
By Melina Morch, Daniel Braun
Normalized Low-Rank Adaptation (NoRA) is a lightweight enhancement to the widely used LoRA technique that normalizes the down‑projection matrices during training. By doing so, NoRA stabilizes early optimization dynamics, accelerates convergence, and improves performance across pretraining, supervised fine‑tuning, and reinforcement learning. The method adds no extra trainable parameters or inference‑time cost, making it broadly applicable.
By Jiale Kang, Ziyin Yue, Zheng Zhan, Yangyi Huang, Weiyang Liu
CAST is a critique‑aware training framework that transforms sparse task outcomes into action‑level supervision for both critique learning and policy optimization. By analyzing agent trajectories, CAST synthesizes structured rationales that explain action validity under partial observability, enabling the creation of richer training data. Fine‑tuned Qwen3‑family models trained with CAST show significant reliability gains, outperforming GPT‑OSS‑120B by over 10% on Retail tasks and improving Telehealth performance by 9% in an out‑of‑domain setting.
By Amir Saeidi, Zehua Zhang, Rishitosh Singh, Naman Ahuja, Vivek Gupta, Ali Payani, Gaowen Liu, Jayanth Srinivasa, Chitta Baral
The paper introduces ClinMPO, an evidence‑guided reinforcement‑learning framework that enhances psychiatric reasoning in small language models (SLMs). ClinMPO leverages a reward model (ClinRM) trained on 18,569 question–answer pairs from 4,474 psychiatry articles and is guided by the psychiatrist‑defined Clinical Psychiatry Thinking Strategy (CPTS). Evaluations on four Qwen3 model sizes show that ClinMPO outperforms baseline, supervised fine‑tuning, and standard policy optimization, with the 8B model surpassing a human baseline of senior pre‑licensure medical students and achieving the highest rank among 31 models. Blinded clinician assessment confirms improved rationale quality across CPTS criteria, demonstrating that clinical evidence and specialist knowledge can be effectively incorporated into medical AI development.
By Xinxin Lin, Guangxin Dai, Yi Zhong, Xiang Li, Xue Xiao, Jian Liu, Yixin Zhang, Lingming Hu, Zhengdong Wu, Yongbo Zheng, Runchuan Zhu, Ming Zhao, Huizi Yu, Yi Zhang, Fangting Lu, Shuo Wu, Jun Zhao, Ping Yin, Joey W. Y. Chan, Ngan Yin Chan, Yumei Wang, Lejin Yang, Yanqiu Xing, Sijing Chen, Yun Kwok Wing, Lin Lu, Xin Ma, Lizhou Fan
The paper introduces a diagnostic framework to disentangle the effects of character profile axes—Familiarity, Structure, and Disposition—on large language model role‑playing agents. Experiments on 211 personas and five LLMs show that Familiarity and Structure have little impact, whereas Disposition, particularly immoral traits, consistently degrades performance. The authors propose Field‑Aware Contrastive Decoding (FACD), a training‑free method that mitigates this performance gap without harming moral‑character performance.
By Yonghyun Jun, Junhyuk Choi, Jeonghyun Park, Jihyeong Park, Liu Nicole Geumheon, Hwanhee Lee
EvoLen is a tokenization method for DNA language models that incorporates evolutionary information to prioritize functional sequence patterns such as regulatory motifs. It groups DNA sequences by cross-species evolutionary signals, trains separate BPE tokenizers for each group, merges vocabularies with a rule that favors preserved patterns, and uses length-aware decoding with dynamic programming. Experiments show EvoLen better preserves functional motifs, differentiates genomic contexts, and aligns with evolutionary constraints while matching or surpassing standard BPE on various DNALM benchmarks.
By Nan Huang, Xiaoxiao Zhou, Junxia Cui, Mario Tapia-Pacheco, Tiffany Amariuta, Yang Li, Jingbo Shang
FRAMEWORKERS is a task‑centric, multi‑agent framework designed for end‑to‑end AI‑generated video production. It uses a central Director to dynamically manage a task stack and an Assistant to execute tasks within a shared Workspace, leveraging modular sub‑agents that can be added without redesigning the workflow. The system is fine‑tuned with supervised learning and policy optimization, outperforming existing LLM planners and fixed pipelines in routing accuracy, failure recovery, and overall video quality.
By Zhendong Li, Lei Sun, Letian Shi, Deheng Zhang, Ruibo Ming, Mengshun Hu, Dannong Xu, Jian Wang, Danda Paudel, Luc Van Gool, Jinjin Gu
arXiv:2608.30051v1 Announce Type: new
Abstract: Process reward models (PRMs) provide dense step-level guidance for search-based reasoning, enabling inference-time compute to be allocated toward promi...
By Taejong Joo, Diego Klabjan
The paper examines how synthetic data can be used to covertly inject targeted social biases into aligned large language models. By having a misaligned teacher model generate filtered synthetic datasets in domains like creative writing and code generation, the authors fine‑tune aligned student models and demonstrate that the benign‑looking data can transmit biases while preserving general task performance. The study highlights a security risk in synthetic data‑driven LLM training pipelines and proposes log‑linearity‑based scoring as a potential safeguard.
By Minkyung Cho, Jihyo Kim, SeungWoo Song, Junghun Yuk, Minjoon Kee, Hoyun Song, KyungTae Lim
The paper presents an agentic retrobiosynthesis framework that uses a learned frontier selection policy to guide multistep retrosynthesis in a deterministic biochemical engine. By fine‑tuning a Qwen2.5‑7B policy with LoRA, the authors achieve higher solve rates than traditional MCTS across several benchmarks, demonstrating that route‑supervised frontier selection can improve search efficiency without changing the underlying reaction model.
By Philippe Meyer, Guillaume Gricourt, Thomas Duigou, Joan H\'erisson, Jean-Loup Faulon
Motus2 is a self‑evolving general world model designed for dexterous manipulation. It integrates a shared‑weight model that offers three control interfaces—a policy, a simulator, and an evaluator—forming a closed decision‑and‑learning loop for policy improvement. The system scales both model size and data, progressing from large‑scale monocular egocentric data to synchronized stereo data and robot‑domain adaptation, while also incorporating tactile feedback and a biomimetic platform with dual arms and hands.
By Hongzhe Bi, Zihao Zhou, Yihang Tang, Jingrui Pang, Shuhe Huang, Haitian Liu, Runqing Wang, Shuai Huang, Yichen Wang, Yiming Cheng, Ruowen Zhao, Zhenghua Li, Hengkai Tan, Xiaolong Liu, Jinhui Wan, Jiabao Liu, Min Zhao, Fan Bao, Jun Zhu