SABLE is a Bayesian optimisation method designed for high‑dimensional latent spaces where the latent dimension far exceeds the evaluation budget. It uses a Bayesian affine surrogate that decomposes each query into an in‑span component optimized by expected improvement and an off‑span component governed by a scalar norm, with a finite‑horizon schedule to avoid under‑exploration. A recompression scheme reduces the number of pseudo‑observations to ≠O(M), keeping refitting costs manageable while achieving high sample efficiency and rapid decision times on synthetic and latent‑space optimisation tasks.
By Erik Bodin, Lisa Maria Kreusser, Donney Fan, Henry Moss, Geoff Pleiss, Nicolas Durrande
CausalCoT is an inference framework that enhances causal transparency in chain-of-thought prompting by structuring problems as a directed acyclic graph of atomic sub-questions. Each node is evaluated in an isolated context that includes only its direct causal parents, thereby preventing attention leakage to unrelated content. This approach maintains competitive predictive performance while significantly improving explanation faithfulness, raising causal sensitivity of intermediate reasoning steps by an average of 12 percentage points and up to 21 percentage points over the strongest baseline, and enables predictable test-time interventions through the structured dependency graph.
By Gabriele Dominici, Teague McMillan, Fazl Barez, Marc Langheinrich, Martin Gjoreski
The paper introduces baseVultra, a compact recurrent neural network cell that reallocates most of its recurrent capacity to a single dense candidate‑generation matrix while simplifying gate operations to O(H) diagonal interactions. On character‑level language modeling benchmarks such as TinyShakespeare and text8, baseVultra matches or slightly outperforms parameter‑matched LSTM and GRU cells, achieving lower validation loss and competitive bits‑per‑character scores. The authors find that the advantages of this candidate‑focused design are most pronounced in low‑capacity regimes, whereas full‑gate cells remain competitive when hidden size is increased.
By Linke Zhang
EgoHand is a monocular 3D hand reconstruction model that sets new state‑of‑the‑art results on egocentric video. It is trained on a curated 14‑million‑image corpus that blends synthetic data, in‑the‑wild video, and multi‑camera motion capture to balance diversity and supervision quality. The model uses SOMA‑Hand as its parametric representation, processes 2D hand crops with a transformer encoder, and incorporates camera cues and an auxiliary 2D branch to enhance geometric consistency and image alignment, achieving low inference latency and strong out‑of‑domain generalization on benchmarks such as HOT3D, ARCTIC, and H2O.
By Chao-Yeh Chen, Tiffany Chen, Lin Duan, Tianyi Hu, Yu-Hsiang Huang, Umar Iqbal, Jiefeng Li, Edy Lim, Yanchen Liu, Yihan Pang, Hang Qiu, Jun Saito, Shao Su, Qingsi Wang, Xiufeng Xie
Route-and-Generate (Rouge) is a hierarchical training framework that lets large language models decide how much reasoning to apply to each input. It separates reasoning effort from answer generation by using a routing policy that distributes a fixed sampling budget across different reasoning formats, while the base model is trained with GRPO and per-sample rewards. The method employs a rank‑based reward favoring shorter correct solutions and updates the router through a closed‑form KL‑regularized policy improvement, leading to more concise answers on easy tasks and longer reasoning on harder ones, thereby improving accuracy‑efficiency trade‑offs on multiple reasoning benchmarks.
By Tien-Phat Nguyen, Truong Nguyen, Thanh-Thien Le, Hoang Phan, Thien Huu Nguyen, Linh Ngo Van, Trung Le
The paper proposes a reinforcement learning approach that rewards language models for producing high‑quality moral arguments, under the hypothesis that better arguing leads to better judging. Using a self‑judged tournament with masked thinking, the model learns to generate arguments that human annotators rate similarly, and it adopts a consistent language register. Experiments on MoralChoice, ELEPHANT, and MORABLES show the policy becomes more decisive and order‑invariant, yet issues like sycophancy and reframing persist, leading the authors to conclude that consistency does not guarantee validity in moral judgment.
By Marius De Kuthy Meurers, Aaron Parisi, Bernd Bohnet
The paper introduces BELT (BELief-based Termination), a method that decouples outcome prediction from stopping control in large language models (LLMs). BELT learns beliefs about current and future answers from a frozen LLM’s hidden states and applies a linear stopping policy, enabling a single probe to adapt to various token‑price trade‑offs without retraining. Across five out‑of‑distribution math benchmarks, BELT cuts token usage by up to 51% while maintaining accuracy, and its lightweight probe outperforms stronger baselines in both efficiency and parameter count, also transferring effectively to coding and agentic tasks.
By Jingwei Ni, Dmitry Abulkhanov, Vladislav Smirnov, Markus Leippold, Elliott Ash, Timothy Baldwin, Artem Shelmanov, Mrinmaya Sachan
The paper presents a scalable method for learning dataweights that optimizes training-example importance using metagradients, guided by a small high‑quality target dataset. It introduces parametric weighting rules that generalize to new and streaming data and can transfer from smaller to larger models. Applied to reward modeling, the learned weights improve held‑out performance, adapt models to specialized domains, and enable human‑synthetic supervision mixtures to outperform either source alone.
By Hedi Driss, Evan Frick, Anastasios N. Angelopoulos, Cho-Jui Hsieh, Yejin Huh, Tijana Zrnic, Andrew Ilyas
The paper introduces WholeBodyMem, an object-guided perceptual memory for humanoid loco‑manipulation that records key frames of object events before the robot acts on its goal. Using SAM3 to track task objects, the system selects causally relevant frames within an adaptive budget and feeds them to an action expert via cross‑attention, along with tokens encoding appearance, geometry, and time. The authors also present HuMeGo, a benchmark with simulated and real‑world Unitree G1 tasks, showing that WholeBodyMem significantly improves success rates over memoryless and uniform‑memory baselines—raising policy success from 29.1% to 72.6% in simulation and from 36.7% to 88.3% on the real robot, while also boosting standard SIMPLE tasks from 70.2% to 86.9%.
By Chenyuan Chen, Ran Ding, Lechen Zhang, Zewen He, Yoshihiko Nakamura
The study investigates how different prompting strategies affect large language model (LLM) judges across multiple languages and tasks. Using a benchmark of 9 prompting strategies, 4 open-weight judges, 27 languages, and 4 tasks, the authors collected over 1 million pointwise scores validated against human judgments. Results show that prompting effectiveness is highly dependent on the judge model, can reverse direction across judges, and becomes less stable in low‑resource languages, indicating that prompt choice must be tailored to the specific evaluation setting.
By Asal Mehrabi, Mohammad Aliannejadi, Clemencia Siro
Sparse Equation Matching (SEM) is a derivative‑free learning framework that uses integral‑based sparse regression with Green’s functions to discover differential operators and driving functions in general‑order dynamical systems. The method unifies several existing equation discovery techniques and is validated through extensive simulations that compare its performance to derivative‑based approaches. SEM is then applied to EEG data from 52 participants performing oculomotor tasks, successfully identifying active brain regions and task‑specific connectivity patterns.
By Jiaqiang Li, Jianbin Tan, Xueqin Wang
The paper introduces Brittlebench, a framework and evaluation pipeline that measures how sensitive large language models (LLMs) are to prompt variations. By applying semantics‑preserving perturbations to popular benchmarks, the authors show that model performance can drop by up to 12% and that such perturbations can change model rankings in 63% of cases. The study finds that these input variations can explain up to half of the performance variance for a given model, underscoring the importance of robust evaluation methods.
By Angelika Romanou, Mark Ibrahim, Candace Ross, Kerem Oktar, Chantal Shaib, Samuel J. Bell, Anaelia Ovalle, Antoine Bosselut, Jesse Dodge, Koustuv Sinha, Adina Williams
The paper introduces simplex denoising, a generative framework that operates on the probability simplex rather than directly in the discrete state space. It employs a non‑Markovian noising scheme where noisy representations at different times are conditionally independent, preserving theoretical guarantees while removing unnecessary constraints. Empirical results show that unrestrained simplex denoising outperforms strong discrete diffusion and flow‑matching baselines on both synthetic and real‑world graph benchmarks.
By Yoann Boget, Alexandros Kalousis
The paper introduces the Graph Memory Transformer (GMT), a decoder‑only language model that replaces the traditional Feed‑Forward Network sublayer with an explicit learned memory graph. In GMT, token representations are routed through a bank of centroids linked by a directed transition matrix, using gravitational source routing and token‑conditioned target selection to produce gated displacements instead of retrieved values. The resulting model contains 82.2 M parameters, trains stably, and offers direct inspection of centroid usage and transition structure, though it slightly lags a larger GPT‑style baseline in validation loss and perplexity.
By Nicola Zanarini, Niccol\`o Ferrari, Evelina Lamma
The paper introduces SAEParate, a sparse autoencoder approach that explicitly separates overlapping concepts in the latent space of text-to-image diffusion models. It employs concept-aware contrastive learning and cross-attention-guided spatial aggregation to focus on concept-relevant regions, while a nonlinear encoder transformation enhances expressive capacity. Experiments on UnlearnCanvas show state‑of‑the‑art performance, especially in challenging joint style‑object unlearning scenarios where prior methods struggle with interference between concepts.
By Hyeonjin Kim, Hangyeol Jung, Heechan Yun, Sungjun Yun, Dong-Jun Han
The paper introduces MASEM, a method for uniform sampling on manifolds that may have multiple disconnected components defined by smooth constraints. It employs a resampling scheme that maximizes the entropy of the empirical distribution using k‑nearest neighbor density estimation, and shows that this approach reduces KL‑divergence exponentially with the number of resampling steps. Experiments on synthetic and robotics benchmarks demonstrate that MASEM achieves faster mixing and improves Sinkhorn distance by an order of magnitude compared to existing alternatives while maintaining competitive runtime.
By Cornelius V. Braun, Tilman Burghoff, Marc Toussaint
SoundnessBench is a benchmark of 1,099 real research proposals designed to test whether large language models can judge the methodological viability of a research idea before resources are invested. The study evaluates 12 frontier LLMs and finds a pervasive optimism bias, with standard prompting yielding a 74.0% false‑positive rate and aggressive prompting increasing false rejections. Analyses show that this bias persists even when controlling for public corpus contamination, surface features, and literature grounding, indicating that current LLMs are not yet reliable gatekeepers for scientific rigor.
By Sy-Tuyen Ho, Minghui Liu, Huy Nghiem, Furong Huang
The paper introduces Riemannian Hodge Message Passing (RHMP), a neural network architecture that separates topology from geometry in physical field modeling on meshes. RHMP fixes the cellular coboundaries determined by mesh incidence and learns symmetric positive‑definite cochain metrics to guide geometry‑dependent propagation, ensuring exact cochain‑complex identities and cochain‑frame equivariance. Experiments on seven benchmarks—including fluids, electromagnetism, gauge fields, and variable‑mesh CFD—show that RHMP achieves strong performance, especially when topology, learned geometry, and field structure interact.
By Dongzhe Zheng, Christine Allen-Blanchette
The paper investigates clustering of multi‑layer networks generated under a mixture multi‑layer stochastic block model (MMSBM). It derives the minimax optimal clustering error rate, expressed via Chernoff information between component edge‑probability distributions, and introduces a two‑stage estimator that achieves this rate even under extreme sparsity. The method is extended to count‑weighted networks using a mixture multi‑layer Poisson block model, with numerical experiments and real‑data analyses demonstrating its practical performance.
By Zhongyuan Lyu, Ting Li, Dong Xia
The paper investigates how to add new skills to large multimodal models (LMMs) without erasing previously learned abilities. By sequentially fine‑tuning on five target skills and tracking performance on eight held‑out benchmarks across three model families, the authors discover that forgetting on one skill can partially recover when the model is later tuned on a different skill. They attribute this to shifts in the output token distribution and propose two simple, selective tuning recipes—updating only self‑attention projection layers or only the MLP Gate&Up while freezing the Down projection—that achieve a better learning‑forgiveness trade‑off than full‑LLM tuning or other common forgetting‑mitigation methods.
whyItMatters":"The study shows that carefully choosing which model components to fine‑tune can enable LMMs to acquire new skills while preserving prior knowledge, offering a simpler and more effective alternative to existing forgetting‑mitigation techniques."
By Zhen Zhu, Yiming Gong, Yao Xiao, Yaoyao Liu, Derek Hoiem