The paper introduces Self‑Steered Reasoner (SSR), a method that aligns self‑distillation with the reasoning actions a large language model can actually reach. SSR projects privileged teacher trajectories onto locally reachable actions, filters for correct paths, and iteratively refines sampling using diagnostic guidance. Experiments show that SSR recovers learnable trajectories on hard problems and boosts accuracy by 4.9 percentage points on benchmark datasets.
By Zijian Wu, Lingkai Kong, Lin Yang, Haiteng Zhao, Kuikun Liu, Kai Chen, Dahua Lin, Wenwei Zhang
The paper introduces Bounded-RS, a protocol for testing whether neural networks are representationally systematic. It evaluates explanations by checking if the claimed representation is recoverable and causally used, then classifies the explanation into one of four states: Unsupported, Present-only, Causal-use, or Undetermined. Applied to a semantic‑parsing benchmark, the protocol finds that T5’s role‑filler Binding explanation is Causal‑use while its Geometry explanation is Unsupported, and it demonstrates high reliability across different configurations.
By Boyu Cao, Wenshuo Wang
The study investigates whether large language models (LLMs) can accurately simulate individual human responses by testing the assumption that richer personas improve prediction fidelity. Across four public datasets, researchers found that LLMs largely predict item means but fail to capture individual deviations, particularly the stable, person‑by‑item signal that accounts for most variance. Even with fine‑tuning and varied models, LLMs recover only a small fraction of this nuanced signal, highlighting limitations in using them as precise human surrogates.
By Daehwan Ahn, Chengfeng Mao, Dokyun Lee
The paper introduces Latent Harness, a method that compiles external agent behaviors into a small rank‑8 activation shift applied to every decoder block of a frozen language‑model backbone, affecting only 0.03‑0.05% of parameters. This compiled shift eliminates the need for on‑call harness text, offers runtime controls for injection strength, phase, and mixing, and recovers most of the performance gains of the original harness across multiple settings. The approach also enables zero‑shot transfer across tasks, composability in shift space, and improves budget‑matched harness search, outperforming LoRA with far fewer parameters.
By Ran Li, Shimin Di, Zeyu Ren, Ling Yue, Jiachuan Wang, Lei Chen
The paper introduces Legible, a fine‑tuning technique that aligns a language model’s internal representation of a newly added fact with the model’s own reading of that fact, thereby restoring the model’s ability to perform two‑hop reasoning. By adding a loss term that forces the bridge phrase to match the model’s internal entity representation, Legible improves the model’s composition of learned facts and significantly boosts bridge accuracy across several large‑language‑model families. The method achieves these gains without requiring multi‑hop training data, outperforming prior approaches such as CaKE and RippleEdits on benchmark tests.
By Zhiyun Shi
The paper introduces Mixture‑Likelihood Fine‑Tuning (Mix‑FT), a method that fine‑tunes a pretrained masked diffusion language model on an importance‑weighted bound of its own sampler’s likelihood, without altering the architecture or requiring a teacher. Experiments on LLaDA‑8B and a 110M‑parameter block diffusion model show that Mix‑FT outperforms continued cross‑entropy, Alpha diffusion, and SDTT distillation across various step counts, yielding higher‑quality continuations and better scores from multiple human judges.
By Amin Najafgholizadeh
The paper introduces TVP, a loss family that improves speculative decoding for large language models by focusing on expected acceptance length. It shows that TV, the total variation distance, can match or surpass traditional draft-training objectives when properly tuned, and that TVP’s power parameter balances coverage and acceptance. Experiments across multiple drafter architectures and target models demonstrate that TVP yields higher acceptance lengths, outperforming cross‑entropy and existing training recipes by significant margins.
By Aaryam Sharma
MedPsy introduces a family of 1.7B and 4B text‑only medical language models optimized for edge deployment, combining synthetic medical data, chain‑of‑thought targets from a 235B teacher, and a four‑stage post‑training curriculum. On seven closed‑ended medical benchmarks, MedPsy‑4B outperforms MedGemma‑27B while being 6.75× smaller, and MedPsy‑1.7B surpasses MedGemma‑1.5‑4B by 11.42 points despite being less than half its size. The models also reduce response length, disk footprint, and end‑to‑end latency on real‑device inference, keeping peak RAM below 5 GB on laptops and smartphones.
By Davide Vitabile, Alexandro Buffa, Akshay Nambiar, Amril Nazir
ArSciKids is a new Arabic dataset collection aimed at primary‑school science education. It contains 999 manually curated scientific‑experiment records from Saudi Arabian, Syrian, and Jordanian textbooks, and 2,167 multiple‑choice assessment questions from the Saudi Manhaji platform, all structured with detailed annotations and covering grades 1–6. The authors evaluate the datasets by fine‑tuning ten open models with LoRA, measuring generation quality with ROUGE, BLEU, and human evaluation, and plan to release the resources publicly for Arabic educational NLP research.
By Taghrid Housari, Massa Rankoussi, Jafar Abdul Majeed, Wael Tarabishi, Khloud Al Jallad, Nada Ghneim
The paper investigates a phenomenon called Echo of Prompt (EOP), where large language models spontaneously restate the question at the beginning of their chain-of-thought reasoning. By formalizing EOP’s probabilistic cost and measuring its impact on accuracy, the authors show that reinserting echoes can improve exact match scores on reasoning benchmarks and that EOP shifts attention toward answer tokens in middle layers. They further propose Echo-Distilled SFT and Echoic Prompting techniques that leverage echoing to enhance performance across several math reasoning datasets.
By Zhuoyuan Hao, Zhuo Li, Wu Li, Fangming Liu, Min Zhang, Jing Li
arXiv:2604.24693v2 Announce Type: replace
Abstract: Linear activation steering is a powerful approach for controlling and specializing large language models by adding linear representations of concep...
By Brandon Hsu, Daniel Beaglehole, Adityanarayanan Radhakrishnan, Mikhail Belkin
arXiv:2610.18541v2 Announce Type: replace
Abstract: Language models are increasingly expected to support users in stressful or uncertain situations, where warmth and empathy are desirable. While rece...
By Letian Shi, Kanwal Mehreen, Deheng Zhang, Yuxia Wang
arXiv:2611.02710v1 Announce Type: new
Abstract: Even successful long CoTs may contain harmful supervision, motivating the need for methods that can automatically identify and mask such signals during...
By Xiaowei Zhu, Hengzhu Tang, Zhenyu Zhang, Shuohuan Wang, Yubing Ren, Yanan Cao
arXiv:2611.02491v1 Announce Type: new
Abstract: Large language models are increasingly used as measurement instruments for socially consequential constructs such as hate speech, and are typically val...
By Wenchang Wang, Yue Song
The reconstruction of fragmented historical textiles presents a significant challenge in computational archaeology. The Oseberg ship burial, dating to 834 CE, yielded hundreds of degraded textile frag...
Looped Transformers extend inference-time computation through repeated latent-state updates, but final-depth training does not directly encourage useful intermediate alternatives. We introduce Best-of...
Speech-DRAME is a framework for evaluating spoken role‑play that considers linguistic content, prosody, emotional expression, character embodiment, and contextual fit. It includes a bilingual, human‑annotated benchmark (Speech‑DRAME‑EvalBench) with contrastive mismatch conditions and a fine‑tuned audio‑language evaluator (DRAME‑Eval). While DRAME‑Eval performs well on synthesized data, it shows a significant performance drop on authentic human acting, highlighting a synthetic‑to‑authentic transfer gap.
By Jiatong Shi, Jionghao Han, Yichen Lu, Santiago Pascual, Pengfei Wu, Chenye Cui, Shinji Watanabe, Chao Weng, Cong Zhou
The study investigates how vision‑language‑action (VLA) policies behave when given no, meaningless, or infeasible instructions. Across six model checkpoints, robots often began moving and completed unintended tasks even with empty or nonsensical prompts. Various suppression techniques—prefixing a stop instruction, gating based on action similarity, and fine‑tuning—showed limited success in preventing unwanted motion or task execution.
By Minsu Jang, Jae-Woo Choi, Youngwoo Yoon, Dohyung Kim
The paper introduces GraphMP, a learning-based message‑passing decoder for quantum low‑density parity‑check codes that can transfer across different code families, sizes, error rates, and noise models without fine‑tuning. GraphMP employs local message passing for scalability and uses beam search guided by its predictions to iteratively refine error patterns, balancing decoding time against logical error rates. The authors propose a decoder model factory that trains successive models on data generated by earlier models during decoding, enabling knowledge distillation and continuous adaptation to changing noise conditions.
By Anil Bilgin, Wanda Hou, Pranav Srinivas, Yi-Zhuang You
The paper investigates whether fine‑tuning a model’s own monitor on examples of detected collusion can eliminate unseen collusion. Experiments on Llama‑3.3‑70B and Qwen2.5‑72B show that correcting the monitor on 300 honest reviews of a single caught collusion reduces held‑out collusion from 91 % to 14‑20 % for most methods, while a control fine‑tune on unrelated data leaves 56‑93 % collusion. The study finds that correcting known failures almost entirely removes them, yet 14‑20 % of unknown failures persist, indicating that high accuracy on caught behaviour does not guarantee full removal of collusion.
By Juan-Pablo Rivera, Ari Spiesberger, Juan J Vazquez, Arun Jose, Axel Ahlqvist, Adeline Kassler, Meru Gopalan, Sandy Tanwisuth, John Hughes