https://techgazette.com.br/research/#learning This week's papers in Machine learning, each with what its authors did and found, in sentences from its own abstract. en 180 Fri, 16 Oct 2026 12:00:00 GMT https://dl.acm.org/doi/10.1145/3859244 https://dl.acm.org/doi/10.1145/3859244 Fri, 16 Oct 2026 12:00:00 GMT ACM TKDD (Knowledge Discovery from Data) Machine learning ACM TKDD (Knowledge Discovery from Data) — Did: Numerous survey studies have covered various aspects of LLMs, such as learning strategies, applications, alignment, explainability, and evaluation. https://dl.acm.org/doi/10.1145/3849700 https://dl.acm.org/doi/10.1145/3849700 Fri, 16 Oct 2026 12:00:00 GMT ACM TKDD (Knowledge Discovery from Data) Machine learning ACM TKDD (Knowledge Discovery from Data) — Did: Adapting large language models (LLMs) for time series forecasting introduces unique challenges due to the mismatch between continuous time series tokens and the discrete word tokens LLMs are trained on. https://arxiv.org/abs/2610.22358 https://arxiv.org/abs/2610.22358 Thu, 15 Oct 2026 17:58:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: To model this dependence, we introduce PathFM, a trajectory-based forecasting model that uses flow matching to generate complete future paths. Found: Experiments on chaotic systems further show that our flow model sustains long-run variability, while strong baselines rapidly collapse. By Jiachen Yao, Xiyuan Zhang, Annan Yu, Boran Han, Abdul Fatir Ansari, Shuai Zhang. https://arxiv.org/abs/2610.22353 https://arxiv.org/abs/2610.22353 Thu, 15 Oct 2026 17:58:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We identify an exploration dilemma between these designs: shared experts can make routing less exploratory by reducing the variability of token representations. Found: MirrorMoE consistently improves performance across model scales. By Bingcong Li, Hao Ma, Michael Muehlebach, Niao He, Yi Ma. https://arxiv.org/abs/2610.22339 https://arxiv.org/abs/2610.22339 Thu, 15 Oct 2026 17:55:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: To this end, we propose LeanGraphSearch, a graph-based retrieval method for Lean 4 that constructs a graph from Mathlib's forward dependencies and traverses it to uncover supporting lemmas through multi-hop dependency chains. Found: Across three theorem proving benchmarks, LeanGraphSearch improves pass rate at {8, 16, 32} reflection turns by as much as 40% compared to vanilla prover and at most 10% compared to LeanSearch v2, showing higher accuracy-cost efficiency. By Yihan Xu, Jia-Chen Gu, Junyi Zhang, Di Wu, Haikang Deng, Po-Nien Kung. https://arxiv.org/abs/2610.22336 https://arxiv.org/abs/2610.22336 Thu, 15 Oct 2026 17:54:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We develop a robust reinforcement learning (RL) framework for learning adaptive treatment policies that generalize across patients and remain reliable under errors in the measured biomarker and shifts in the disease dynamics. Found: Our findings suggest that robust RL holds great promise as a principled methodology for learning reliable adaptive treatment policies. By Chengze Xie, Xuan Wang, Ying Ding, Hairong Wang, Yan Li. https://arxiv.org/abs/2610.22327 https://arxiv.org/abs/2610.22327 Thu, 15 Oct 2026 17:52:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We instead introduce complementary representations as a design principle for time-series reasoning. Found: Together, these results demonstrate that complementary representations provide an effective design principle for general-purpose TSLMs, enabling models to exploit temporal structure without sacrificing access to the original measurement frame. By Hedi Zisling, Liran Nochumsohn, Ilan Naiman, Mihir Parmar, Omri Azencot. https://arxiv.org/abs/2610.22316 https://arxiv.org/abs/2610.22316 Thu, 15 Oct 2026 17:51:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: In this work, we propose a steering agent that supervises execution on the fly and steers the system back on track when drift occurs. Found: Experiments on information-seeking search tasks show that the steering agent provides effective supervision for both single-agent and multi-agent systems, improving their performance and efficiency. By Leyang Shen, Yaxin Gong, Yang Zhang, Chenxiao Fan, Jiakai Wang, Dong Wang. https://arxiv.org/abs/2610.22293 https://arxiv.org/abs/2610.22293 Thu, 15 Oct 2026 17:44:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: In this work, we introduce Subspace Tuning (SST), which restricts each weight update to the bottom singular subspace, eliminating changes along the principal directions. Found: With these mechanisms, protecting as much as 99\% of the weight subspace substantially improves retention while preserving strong target-task performance. By Arnav Chavan, Kusha Sareen, Rishabh Tiwari, Abhranil Chandra, Johan Obando-Ceron, Kurt Keutzer. https://arxiv.org/abs/2610.22292 https://arxiv.org/abs/2610.22292 Thu, 15 Oct 2026 17:43:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We introduce Continuous Adversarial Training for Consistent Honesty (CATCH), inspired by a method originally developed to robustly refuse harmful queries. Found: However, CATCH surpasses honesty steering in maintaining utility, thanks to a novel choice of utility data. By Yuan-Chen Chang, Niklas Herbster, Gauthier Gidel, Tommaso Tosato. https://arxiv.org/abs/2610.22287 https://arxiv.org/abs/2610.22287 Thu, 15 Oct 2026 17:42:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We study rational synthesis for stochastic multi-agent systems with unknown dynamics, where each agent is equipped with its own probabilistic temporal objective specified using Probabilistic Linear Temporal Logic on Finite Traces ($\mathsf{PLTL}_f}$). By Omar Adalat, Edwin Hamel-De le Court, Francesco Belardinelli. https://arxiv.org/abs/2610.22273 https://arxiv.org/abs/2610.22273 Thu, 15 Oct 2026 17:35:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We build on a key asymmetry in modern agentic harnesses (e.g., Claude Code): while LLM agents are adept at navigating existing filesystems to solve an immediate task, they are rarely trained to construct reusable workspaces for future use. Found: Empirically, Librarian Agents outperforms the strongest retrieval and precomputation baselines by 33 percentage points on BrowseComp-Plus, amortizes upfront construction costs across repeated queries, and generalizes zero-shot to unseen domains on BRIGHT with a 31-point average accuracy improvement. By Anikait Singh, Teresa Zhang, Yoonho Lee, Roshen Sanjay Nair, Aviral Kumar, Chelsea Finn. https://arxiv.org/abs/2610.22263 https://arxiv.org/abs/2610.22263 Thu, 15 Oct 2026 17:28:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: Reinforcement learning has become a widely used approach to post-training large language models. By Shuo Han, Ziqing Hu. https://arxiv.org/abs/2610.22262 https://arxiv.org/abs/2610.22262 Thu, 15 Oct 2026 17:28:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: Scientific super-resolution (SR) models are typically optimized with pointwise reconstruction losses, yet scientific interpretation often hinges on extrema, connected regions, and multiscale scalar structure. Found: Mean MT distance also improves, but on only 91 of those 150 fields, indicating that PD and MT capture complementary structure. By Arjun Dadhwal, Brian Summa. https://arxiv.org/abs/2610.22256 https://arxiv.org/abs/2610.22256 Thu, 15 Oct 2026 17:24:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We present BitSparse, a GPU-resident lossless compression format for sparse activations. By Max Zhu, Bohang Sun, Pietro Lio, Jialin Yu. https://arxiv.org/abs/2610.22252 https://arxiv.org/abs/2610.22252 Thu, 15 Oct 2026 17:22:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We study how to evaluate $K$ target policies simultaneously in stochastic contextual bandits with linear reward structure. By Spyros Dragazis, Yilei Chen, Aldo Pacchiano. https://arxiv.org/abs/2610.22242 https://arxiv.org/abs/2610.22242 Thu, 15 Oct 2026 17:16:00 GMT arXiv stat.ML Machine learning arXiv stat.ML — Did: We address this issue by introducing SurvICL, a TFM that predicts survival curves in-context. Found: On the 63 datasets held out from all design decisions, SurvICL outperforms 20 tree-based, deep, and foundation-model baselines without any tuning: it wins 82.9\% of pairwise comparisons in integrated Brier score, against 71.3\% for the runner-up, the Cox model, and is the first model to rank above Cox on small datasets. By Julie Alberge, Linus Bleistein, Judith Abécassis, Marine Le Morvan, Gaël Varoquaux, David Holzmüller. https://arxiv.org/abs/2610.22238 https://arxiv.org/abs/2610.22238 Thu, 15 Oct 2026 17:15:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We propose the Single-Increment Rejection Sampler (SIRS), which provides an estimate of the log acceptance probability from a single randomly sampled increment at the cost of one query. Found: We show that the KL-divergence between the data distribution and the SIRS output is bounded by a clipping term plus the model's training error. By Shiyuan Zhang, Xuheng Li, Qiwei Di, Quanquan Gu. https://arxiv.org/abs/2610.22236 https://arxiv.org/abs/2610.22236 Thu, 15 Oct 2026 17:15:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We show that the consolidation does not need training: with forward passes only, the capabilities of separately tuned specialists can be aggregated into one model that is, on average, more accurate than either specialist alone. By Bangji Yang, Jiajun Fan, Zhaoyu Zhu, Hongbo Ma, Ge Liu. https://arxiv.org/abs/2610.22234 https://arxiv.org/abs/2610.22234 Thu, 15 Oct 2026 17:14:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We present OS-Helix, an end-to-end trained multimodal computer-using agent that unifies GUI actions and code execution in one action space. Found: On OSWorld-Verified, OS-Helix-27B and OS-Helix-35B-A3B achieve relative improvements of 35.0% and 33.7% over their matched Qwen3.5 hybrid baselines, while OS-Helix-27B-RL further improves by 42.9% to reach 66.44% success. By Bowen Yang, Kaiming Jin, Zichen Ding, Zhenyu Wu, Zehao Li, Zhoumianze Liu. https://arxiv.org/abs/2610.22213 https://arxiv.org/abs/2610.22213 Thu, 15 Oct 2026 17:00:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: Finally, we develop dimension-independent bounds for loss functions of a specific structure, making the generalization guarantees applicable to high-dimensional problems. By Ruixi Hu, Liang Wu. https://arxiv.org/abs/2610.22179 https://arxiv.org/abs/2610.22179 Thu, 15 Oct 2026 16:34:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: We introduce Losslessly Compressed Tensors (LCT), a GPU execution framework designed for this repeated encode-decode regime. Found: On a 162M-parameter NanoGPT, LCT reduces memory usage by $24.6\%$ for a $7.2\%$ step-time increase. By Max Zhu, Bohang Sun, Pietro Lio, Jialin Yu. https://arxiv.org/abs/2610.22157 https://arxiv.org/abs/2610.22157 Thu, 15 Oct 2026 16:22:00 GMT arXiv cs.LG Machine learning arXiv cs.LG — Did: Despite these advantages, we show that their application to higher-dimensional problems remains limited by the computational cost of training. Found: These results demonstrate that careful consideration of how PI-DeepONets are implemented can yield substantial gains, enhancing the practical performance of PI-DeepONets and enabling their application to otherwise intractable physics-informed learning problems. By Nahum Banks, Gustavo A. M. de Almeida, Sergio Maldonado. https://arxiv.org/abs/2610.21832 https://arxiv.org/abs/2610.21832 Thu, 15 Oct 2026 13:15:00 GMT arXiv (systems & security) cs.LG Machine learning arXiv (systems & security) cs.LG — Did: We prove that this one-stage look-ahead threshold is optimal whenever stopping decisions are nested, derive closed forms for geometric and power-law returns, and show that an estimated rule loses utility only through wrong-signed decisions. By Ke Xiao, Qiyuan Wang, Christos Anagnostopoulos. https://arxiv.org/abs/2610.21640 https://arxiv.org/abs/2610.21640 Thu, 15 Oct 2026 10:21:00 GMT arXiv (systems & security) cs.LG Machine learning arXiv (systems & security) cs.LG — Did: This paper presents ReFlow, a system-algorithm co-designed RL framework for efficient LLM post-training. Found: Experiments across representative single-turn reasoning and multi-turn agentic RL tasks show that ReFlow improves training throughput by up to $2.06\times$ over state-of-the-art synchronous and asynchronous RL systems, while maintaining comparable or better model performance. By Yukun Tian, Junwei Su, Haofeng Xu, Lansong Diao, Zhengping Qian, Weibo Zhao.