The paper introduces Prototype-Mediated Process Supervision (PMPS), a method that uses learnable reasoning prototypes to provide structural supervision for latent chain-of-thought embeddings, addressing representation collapse and uneven information distribution. PMPS aligns latent and explicit CoT embeddings in a shared prototype space via many-to-many soft assignment and employs a Progressive Sequential Alignment module to guide training from positional priors to adaptive matching. Experiments show PMPS reduces output token length to under 50% of explicit CoT on GSM8K-Aug and improves accuracy by 2.08% over SIM-CoT, even surpassing CoT-SFT on GPT-2 and achieving the highest accuracy among latent reasoning methods on larger models and harder tasks.
By Yiqi Li, Xu Chen, Chen Ju, Jiangchao Yao, Zhaoyang Li, Jinsong Lan, Xiaoyong Zhu, Bo Zheng, Yu Wang
The paper introduces SmartWeatherAgent, a three‑stage architecture that combines intent recognition, hazard prediction, and reasoning‑enhanced generation to improve tourism meteorological services. It fuses rule‑based methods with large language models and a LightGBM model enriched with highland‑specific features, achieving an F1‑Macro score of 0.605 and 1.60 ms latency on high‑wind, precipitation, and low‑temperature events. A 12‑round micro‑step prompt self‑optimization loop raises the composite warning quality score from 4.2 to 8.9, with notable gains in data source citation, physical mechanism explanation, and scientific rigor through explicit uncertainty statements.
By Shuai Yan, Yang Xu, Shan He
The paper reports a large‑scale post‑training ternarisation of the Qwen3 language model, extending a conversion pipeline from the 4B to the 8B variant. Using KOTMS rotation, E2M‑ATQ adaptive ternarisation, and GPTQ‑style error compensation, the authors achieve a 1.361× perplexity ratio across three corpora and retain 78.5% of the FP16 accuracy on zero‑shot tasks, with the 8B model outperforming the 4B by 8.9 percentage points. The study also demonstrates lossless lattice‑aware packing, producing an 8.24 GiB checkpoint that preserves perplexity, and shows that direct packed execution can reach 15.52 tokens/s in 7.35 GiB, though packed GEMV remains slower than FP16 cuBLAS.
By Anirudh Malik, M Sparsh Mehra, Poojith Devan
The paper introduces a lightweight evaluation method for logical reasoning in transformer-based language models, using query-key alignments within attention heads to compute a QK-score. This single forward-pass technique identifies valid versus invalid inferences and is validated across multiple reasoning benchmarks, showing robustness to distractors and deeper reasoning. Experiments span models from 1.5B to 70B parameters, demonstrating scalability.
By Eduard Tulchinskii, Anastasia Voznyuk, Laida Kushnareva, Andrei Andriiainen, Irina Piontkovskaya, Evgeny Burnaev, Serguei Barannikov
The study evaluates Vibe Coding, an AI‑led conversational programming paradigm that lets developers generate software via natural‑language interaction with large language models. In a mixed‑methods experiment with 30 participants, Vibe Coding improved development efficiency—reducing task completion time by 27% versus traditional coding and 12% versus AI‑assisted coding—while also yielding a good usability score (SUS = 71.4) and moderate cognitive workload (NASA‑TLX = 55.5). However, the gains came with trade‑offs: lower maintainability indices, higher security vulnerabilities, and themes of trust calibration, loss of control, and prompt‑engineering strategy emerged, leading the authors to propose a three‑pillar framework for responsible adoption.
By Sales G. Aribe Jr., Louie Jay S. Labastida
The paper introduces Weight-Redundancy Pruning (WRP), a forward‑free depth‑pruning technique for large language models that estimates inter‑layer redundancy using only checkpoint weights. WRP compares attention outputs and MLP down‑projection weights across layers, combining pairwise similarities with relative projection‑scale information to guide layer grouping and block selection. Experiments show that WRP consistently outperforms existing forward‑free magnitude pruning methods and approaches the performance of activation‑based pruning across various pruning settings, model families, and downstream tasks.
By Vincent-Daniel Yun, Woosang Lim
CS-Guard is a new benchmark that systematically evaluates guardrails for code generation security, covering 1,000 malware-generation prompts, 7 jailbreak attacks, and a novel fictional scenario attack (FSA) for text-to-code generation, as well as 331 code prompts for code-to-code generation. The study empirically tests nine guardrails across seven large language models, finding that many guardrails fail to prevent malicious code generation, with attack success rates reaching about 50% for text-to-code and up to nearly 100% for code-to-code and FSA scenarios. CS-Guard introduces a modular three-layer guardrail taxonomy and releases its benchmark and data to support future research.
By Jinyang Li, Mingyu Guo, Hung X. Nguyen
The paper introduces Forgetting Only What Matters via Unlearning Layers (FOM-UL), a layer‑selective unlearning framework for large language models. FOM-UL uses a forget‑to‑retain significance score to identify transformer layers that strongly influence the forget set while being insensitive to the retain set, allowing targeted updates that preserve most of the model. Experiments on TOFU, KnowUnDo, and MUSE-style benchmarks show that FOM-UL reduces residual memorization and maintains utility better than several baselines, even after 8‑bit and 4‑bit post‑training quantization, and it also limits recovery of forgotten content in adversarial prompt tests.
By Ravi Ranjan, Olivera Kotevska, Agoritsa Polyzou
Grounded Continuation introduces a runtime verifier that classifies each utterance in an LLM conversation into one of eight epistemic operations and uses a symbolic engine to maintain a dependency map of claims and their supports. The verifier checks whether a new continuation is grounded by walking this map, a linear-time process that requires no additional LLM calls. On benchmarks such as ReviseQA and MemoryAgentBench, the verifier improves single-hop accuracy for several QA models, even enabling a 7B model to outperform GPT‑4o when guided by the verifier.
By Qisong He, Jinwei Hu, Xinmiao Huang, Changshun Wu, Yi Dong, Xiaowei Huang
The paper introduces a Longitudinal Tabular Transformer (LTT) to predict Estimated Times of Restoration (ETRs) for electrical outages by leveraging the sequence of revisions recorded during outage management. Unlike prior static regression approaches, LTT processes each revision as part of a longitudinal sequence, producing refined ETR estimates at every revision. Across 242,928 storm-related outages and 10 million revisions from six utilities, LTT reduces customer-weighted asymmetric error by a median of 36.9% compared to utilities’ published estimates and improves root mean squared error at all companies.
By Bogireddy Sai Prasanna Teja, Valliappan Muthukaruppan, Carls Benjamin
The paper introduces TaMeSo‑bot, a soft‑wrist robotic system that uses tactile memory to perform robust object insertion tasks. It employs a Masked Tactile Trajectory Transformer (MAT³) to jointly model actions, tactile cues, force‑torque data, and proprioception, learning spatiotemporal representations through masked token prediction. Experiments on peg‑in‑hole tasks show that MAT³ outperforms baselines and adapts well to unseen pegs and conditions.
By Tatsuya Kamijo, Mai Nishimura, Nodoka Shibasaki, Jeremy Siburian, Cristian C. Beltran-Hernandez, Masashi Hamaya
Spec‑Harness evaluates how well large language models (LLMs) synthesize Java Modeling Language (JML) specifications by measuring behavioral adequacy across precondition and postcondition correctness and completeness. The study shows that while prompt optimization can raise verifier pass rates, many accepted specifications remain behaviorally weak, either over‑ or under‑constraining inputs and outputs. Spec‑Harness also serves as a feedback mechanism that improves the quality of specifications generated by general‑purpose coding agents and a specialized JML agent.
By Md Rakib Hossain Misu, Iris Ma, Cristina V. Lopes
The study investigates how fine‑tuning large language models on synthetic stories can imprint human character traits onto AI assistants. Even when only a small fraction of stories contain a particular behavior, the assistant adopts that conditional behavior while remaining generally helpful. The researchers find that the assistant is more influenced by characters that resemble its own persona—an effect they call the affinity effect—and that this influence extends to base models and different system prompts.
By Jorio Cocola, Lev McKinney, Harry Mayne, Jan Betley, Owain Evans
The paper investigates how large language models (LLMs) share a common Fisher‑Rao geometry in their next‑token probability distributions, revealing that behaviour largely determines this geometry while activation geometry depends on coordinate choices. Across transformer, state‑space, and recurrent architectures, output geometries align more closely than activation geometries, and this shared structure facilitates semantic‑category transfer and improves agreement with human word choices as models scale and train. The study further demonstrates that geometry can guide minimum‑disturbance interventions, enabling reusable control that preserves behaviour better than Euclidean methods and enhances steering, editing, attribution, dictionary learning, and fine‑tuning.
By Dario Picozzi
The paper introduces SolCloudLLM, a large language model–based framework that fuses sky‑image patches with time‑series data through bidirectional multimodal fusion for short‑term solar forecasting. Experiments on the SIRTA and SKIPP'D datasets show that SolCloudLLM outperforms existing baselines, achieving up to a 25.4% reduction in mean squared error, especially under cloudy conditions and in few‑shot scenarios.
By Ken Chen, Maneesha Perera, Wei Wang, Sachith Seneviratne, Hansani Weeratunge, Saman Halgamuge
The paper presents a few-shot regression framework for plant growth estimation that combines Vision Transformer embeddings, fuzzy c-means clustering for task construction, and gradient-based meta-learning. It demonstrates that structuring tasks in the embedding space is crucial for performance and that second-order meta-learning methods such as MAML++ outperform classical baselines in low-label scenarios. Experiments on two plant datasets confirm that this approach yields reliable growth estimates even with severe label scarcity.
By Sheikh Hasan Elahi, Rusith Chamara Hathurusinghe Dewage, Habib Ullah, Muhammad Salman Siddiqui, Rakibul Islam, Fadi Al Machot
The paper introduces a Kubernetes Dynamic Resource Allocation driver that treats composable CXL memory as a schedulable cluster resource, enabling cross-node shared memory for large language model (LLM) serving. By composing CXL regions on demand, materializing them as DAX devices, and exposing them via a single Container Device Interface name, pods on different nodes can access the same physical memory region. A shared‑memory connector for vLLM/llm‑d uses this region as a KV‑cache tier, eliminating external metadata services and achieving significant reductions in time‑to‑first‑token (TTFT) with minimal additional latency compared to same‑node reuse.
By Hongjian Fan, Kevin Zhang, David Habinsky, Sean Dykstra
The paper introduces Structurally Speaking, a lightweight structured prompting protocol that improves graph captioning by translating explicit connectivity into concise motif-level descriptions. It treats motif-oriented graph captioning as a bidirectional graph-text translation task, aiming to preserve topology for graph recovery while producing interpretable captions. Experiments on a synthetic motif-based dataset show that the protocol yields shorter, more motif-consistent captions without sacrificing graph recovery performance.
By Hsiao-Ying Lu, Dongyu Liu, Kwan-Liu Ma
The paper proposes an information-weighted cross‑entropy loss that rescales token contributions using TF‑IDF statistics, thereby emphasizing semantically informative tokens and down‑weighting ubiquitous ones. Experiments on five decoder‑only language models (1.1B–13B parameters) show consistent reductions in memorized substring length while maintaining perplexity and downstream performance. The method is architecture‑agnostic, adds less than 3% computational overhead, and can be integrated into existing training pipelines.
By Zhijian Li, Stefan Larson, Kevin Leach
CT‑SAFR is a multi‑layered verification framework designed to enhance the safety and faithfulness of Chain‑of‑Thought reasoning in autonomous robots. The framework achieves a 94.2% hallucination detection rate with sub‑500 ms latency, and a warehouse robot case study shows an 87% reduction in unsafe reasoning outputs. The study also offers recommendations for responsible deployment of reasoning‑capable autonomous robots.
By Cagri Temel