The paper introduces MAGS, a multi-agent framework that automatically generates executable programs with formal safety guarantees. MAGS translates LLM-generated code into the verification-aware language Dafny, repairs any safety violations using verifier feedback, and then compiles the verified code back into executable form. Evaluations on 220 diverse examples—including CUDA kernels, terminal scripts, and robotic-arm tasks—show a 100% success rate in producing programs that meet frozen safety specifications, with additional safety and functional tests confirming strong performance across domains.
By Albert Wu, Nicholas Roberts, Tzu-Heng Huang, Haoran Lin, Gil Friedman, Sungjun Cho, Gabriel Orlanski, Frederic Sala
The paper investigates a new failure mode of tool‑augmented large language model agents: calling non‑existent tools with arguments that do not match any declared schema. It introduces a five‑class taxonomy of tool hallucination, presents a training‑free closed‑world resolver that checks registry membership and signatures, and demonstrates that hallucinations persist across ten hosted models and various invocation surfaces, including the Model Context Protocol. The authors release a Hallucinated‑Tools Benchmark to enable comparison of resolver methods.
By Laxmipriya Ganesh Iyer
The paper introduces a dependency‑graph framework to formalize compositional reasoning in language models, defining three increasing levels of compositionality. Using data‑structure tasks with deterministic rewards, the authors observe a consistent asymmetry: training on decomposed skills does not reliably transfer to composed tasks, whereas training on composed tasks transfers more readily to decomposed ones. They provide a theoretical explanation for this asymmetry and evaluate its effects under length extrapolation, structural distribution shift, and transfer to unseen skills, concluding with a pilot study on real‑world tool‑calling benchmarks that suggests the phenomenon extends to practical settings.
By Yu He, Yingxi Li, Yifei Wang, Ellen Vitercik
The paper introduces LLM-as-an-Improver, a method that uses verification feedback to enhance the candidate set in verifier-based selection. It proposes Verify–Repair–Reselect (VRR), which keeps the initial winner, generates three complementary alternatives (repaired versions of the winner and runner‑up, and a new approach), filters invalid or duplicate candidates, and then reselects the final answer. Experiments on code‑generation and reasoning benchmarks show that VRR outperforms fixed‑pool selection and can recover correct solutions even when the initial pool is entirely wrong.
By Akiyoshi Tomihari, Yuma Ichikawa
The paper introduces SIFT, a sample‑efficient framework for self‑improvement of coding agents that uses a fast tree‑search guided by an LLM‑as‑a‑judge signal. By performing pairwise comparisons of candidate patches and aggregating results with a regularized Bradley‑Terry model, SIFT limits expensive downstream evaluations to only the most promising nodes. The approach achieves higher coding performance on the Polyglot benchmark while reducing CPU hours, wall‑clock time, and API costs compared to prior tree‑search self‑evolution methods.
By Xinghong Fu, Aravinth Kulanthaivelu, Yutaro Yamada
The paper introduces LABELSENSE-Pilot, a prototype for placing point-feature labels on interactive maps that balances geometric validity, display yield, local utility, and stability during camera motion. It generates eight compass candidates per feature, scores them with a multilayer perceptron using graph-context summaries, adds a previous-placement bonus, and selects a layout via mixed-integer optimization, ensuring viewport containment, per-feature uniqueness, and pairwise clearance. Experiments on 2,500 airport data show 85.62 % label display with only 2.09 % flicker and no collisions, outperforming a handcrafted integer program in flicker reduction while maintaining comparable display rates.
By Taimoor Ahmad
The study examines how language influences AI chatbot responses to questions about the war in Ukraine, revealing that the same AI systems (GPT, Claude, Gemini) produce varying political stances across 112 languages. By evaluating 20 statements in 112 languages, the researchers found that Russia‑leaning versus Ukraine‑leaning answers differ by language, mirroring global political attitudes such as public support for Russia, UN voting patterns, and aid levels. This pattern persists across all three models and even when specific statement pairs are removed, suggesting that information warfare could embed geopolitical biases into AI training data.
By Maxim Chupilkin
The paper investigates whether gradient‑based data attribution methods can filter out subliminal learning in language models. Three attribution techniques—GradCos, a contrastive variant, and EK‑FAC—are evaluated on three models and compared to divergence tokens, a strong baseline that requires counterfactual teacher models. Results show that EK‑FAC can mitigate some subliminal effects at the token level, but overall the methods are inconsistent and generally less effective than divergence tokens, especially when filtering entire samples.
By Moritz Weckbecker, Sweta Jena, Jonas M\"uller, Ponnurangam Kumaraguru, Sebastian Lapuschkin, Wojciech Samek, Louis Jaburi, Gon\c{c}alo Paulo
The study examined how personalising language models affects user interactions over five days, comparing a non‑personalised baseline with memory‑based and survey‑based personalisation. Results showed that many interaction changes were due to repeated exposure, but personalisation influenced specific behaviors: memory‑based users disclosed more and found the model less creepy, while survey‑based users felt more regret about sharing personal data. The authors emphasize the nuanced, approach‑specific impacts on user attitudes and the need for careful design of personalised AI.
By Canfer Akbulut, Justine Breuch, Arianna Manzini, Lujain Ibrahim, Matija Franklin, Roma Patel, Iason Gabriel, Kristian Lum, Laura Weidinger
The paper introduces FreqCondNorm, a Transformer-based architecture that adds a frequency-conditioned normalization layer to unify heterogeneous time-series data for predictive maintenance. The model is pretrained on five public datasets using masked auto‑encoding and contrastive learning, achieving 99.2% accuracy on CWRU and 82.1% zero‑shot accuracy on MFPT, showing strong transfer across sampling frequencies. However, it does not improve remaining useful life prediction, indicating a mismatch between pretraining and RUL objectives that requires further study.
By Zaynab Raounak, Camille LHermine, Zhiguo Zeng
The paper investigates how agent harnesses—specifically planning guidance, execution organization, and completion verification—affect performance in retail and airline pilot tasks. By comparing fixed, task‑specific plans to shuffled policy text of equal length, the study finds that fixed plans improve success rates by about 7 percentage points, especially on complex tasks. A read‑only verifier rejects a majority of invalid episodes while incurring minimal cost, and its impact varies with the penalty for erroneous acceptance, often matching the full planning‑plus‑verification benefit at a lower cost.
By Yukun Zhang, Kemu Xu, Yishen Chen
Deep Noir is a framework that autonomously discovers optimal activation‑steering parameters in transformer models by leveraging Logit Lens convergence and causal head‑level attribution. It demonstrates significant performance gains across models ranging from 1B to 9B parameters, improving spam detection by up to 42 percentage points and SST‑2 sentiment classification by 13.1 percentage points without code changes. The approach also reveals that increased steering magnitude expands a predictable prompt‑injection attack surface, highlighting security implications for agent systems using steered classifiers.
By Frank E. Bobe III, Gregory D. Vetaw, Darshan W. Bryner, Matthew G. Cook, Jose L. Salas-Vernis
The paper proposes a new framework that improves spreadsheet chunking for large language model (LLM)-driven retrieval-augmented generation (RAG) systems by adding semantic cell annotations. This approach outperforms current state‑of‑the‑art methods but is limited by the inherent two‑dimensional, unstructured nature of spreadsheets, which cannot be fully captured by finite classification categories. The authors argue that future progress requires dimensionality‑reduction techniques to flatten spreadsheets into one‑dimensional text, simplifying downstream RAG interpretation and generation.
By Zofia Smole\'n
The study investigates how individual components of a coding harness—planning, action space, and context management—affect autonomous coding agents’ performance. By fixing the execution loop and varying these components across 176 settings on SWE‑Bench Verified and Terminal‑Bench 2.1, the authors find that context management is most valuable when context windows are tight, staging rule‑based elision before LLM summarization yields the best efficiency, planning serves as an accuracy scaffold for weaker models and a cost saver for stronger ones, and predefined tools help models with limited bash skills while bash‑capable models benefit from a bash‑only interface. Trajectory‑level analysis shows that context management lengthens execution paths, planning alters where trajectories terminate, and the action space determines code granularity, offering a modular framework for future harness design.
By Run-Ze Fan, Zihao Zhang, Simin Ma, Yebowen Hu, Shouju Wang, Kaiqiang Song, Fei Liu, Hamed Zamani, Xiaoyang Wang
MeshKV is a Network‑on‑Chip key‑value cache fabric designed to improve transformer decoding on tiled accelerators. It uses packetized flows, TaKV affine striping, Mare multicast with duplicate suppression, and Pad to overlap prefetch, multiply, and softmax operations, thereby converting back‑pressure into useful KV transfer. In an 8×8 FPGA prototype with LLaMA‑2‑7B and Mistral‑7B, MeshKV cuts interconnect traffic by up to 58 %, doubles KV bandwidth utilization, and boosts multi‑stream throughput by up to 1.9×.
By Dong Liu, Yanxuan Yu
The paper proposes a layer-wise curriculum learning strategy for compressing large language models (LLMs). By partitioning the model into layer segments and starting training with easier tasks before progressing to harder ones, the method accelerates convergence and stabilizes knowledge transfer from teacher to student models. Additional techniques such as feature caching with multi-threading improve GPU utilization, leading to state‑of‑the‑art compression results and over 50% reductions in memory usage and training time on BERT and GPT‑2, while outperforming other pruning methods on LLaMA‑family and Qwen models.
By Donggeon Lee, Dooyeon Na, Seungmin Oh, Jongbin Ryu
The YNU-HPCC team participated in Subtask A of SemEval‑2025 Task 11, "Bridging the Gap in Text‑Based Emotion," using a RoBERTa model with a single prediction head to process one emotion at a time. Their system achieved an official ranking score of 0.44 across all languages after translating the dataset into English with Google Translate. Analysis showed that a single head outperformed six simultaneous heads and that training on the uniformly translated English data improved results.
By Hao Yang, Jin Wang, Xuejie Zhang
The paper investigates why large language models (LLMs) fail to transfer knowledge across languages. By pretraining 360M- and 7B-parameter models, the authors show that poor cross‑lingual generalization arises during pretraining and persists despite standard fixes. Using a controlled bilingual setup with identical text but disjoint token spaces, they demonstrate that token disjointness alone causes knowledge compartmentalization, and that mapping languages into a shared token space via word‑wise translation markedly improves cross‑lingual performance, recovering up to 12.6% of native‑language learning efficiency.
By Adam Gaber, Uriel Dolev, Elisabeth Fittschen, Bobby Cheng, Yuval Marton, Leshem Choshen
The paper investigates whether open‑source Vision‑Language Models (VLMs) can perform zero‑shot action quality assessment (AQA) on Olympic diving videos. Using the AQA‑7 benchmark, the authors propose a regression framework that combines VLM‑generated semantic reasoning, phase‑level sub‑scores, TF‑IDF vectorization, dimensionality reduction, and ensemble learning to predict final competition scores. While individual VLMs achieve moderate Spearman correlations (<0.32), the ensemble approach boosts performance to 0.67, demonstrating that VLM‑derived textual reasoning features are more informative than raw numerical sub‑scores for AQA.
whyItMatters":"The study shows that VLMs can serve as explainable, semi‑automated tools for evaluating sports performance, potentially aiding expert judging in complex, subjective Olympic events."
By Henry O. Velesaca, David Freire-Obregon, Luigi Miranda, Abel Reyes-Angulo
The paper introduces probe guidance, a technique that leverages frozen internal states of a diffusion model to generate a guidance signal without requiring an extra forward pass during inference. This method improves continuous diffusion language models, achieving state‑of‑the‑art results on unconditional generation and enhancing performance on multiple‑choice question answering for a 1.7B model. The authors also use probes to analyze autoguidance, revealing that the weak model must originate from a low‑entropy training region to align dynamics with the strong model.
By Rohit Dilip, Tianrong Chen, Yuyang Wang, David Van Valen, Joshua Susskind, Miguel Angel Bautista