Machine Learning
Kwai Summary Attention V2 Technical Report
OneRec Team, Chenglong Chu, Guohong Mu et al. (+30 more)
Summary-
token architectures such as Kwai Summary
Attention (KSA) compress distant context into learnable Summary
token per chunk, and hybrid variants interleave these layers with full
attention to retain exact access to the original text. In such hybrid KSA-Full architecture, the interleaving inserted summary
tokens are served as a streaming flow role across different layers. Although effective, however, it also introduce some bottlenecks: (1) the different layers summary compression chunk must be same, therefore hard to reach a flexible KV Cache compression, e.g., different layer set different summary compression chunk window. (2) in the full-
attention layers, the summary
token introduced additional computation cost, but such layer ignore them and still pay quadratic prefilling, therefore unfriendly to the computation side. Motivated by them, we developed the KSA V2, which not only achieves a more flexible KV cache
memory compression but also reduces the computation cost in Full
attention layer. Specifically, we found that the summary
tokens can serve two roles: the KV
memory and sparse indexer.
Read the full abstract ▾
All authors OneRec Team, Chenglong Chu, Guohong Mu, Guowang Zhang, Hao Peng, Hui Wang, Jian Liang, Jiangxia Cao, Jiao Ou, Jiaxin Deng, Lu Ren, Muhao Wei, Ruiming Tang, Tianle Xu, Wenwu Ou, Xingmei Wang, Zhixin Ling, Zhiyuan Liang, Ziqi Wang, Chaoyi Ma, Haixiao Gao, Kuo Cai, Renpu Liu, Ruochen Yang, Wenlong Xu, Xinqi Jin, Yan Sun, Yifan Wang, Yingzhi He, Yiyu Wang, Yusen Huo, Zhaojie Liu, Zihan Zhou
Machine Learning · Cross-listed from cs.SD
AMemOps: Evaluating Memory Updates in Long Audio Streams
Peiwei Chang, Li Zhang, Mark D. Plumbley et al. (+2 more)
Audio
memory must recognise recurring sources across acoustic changes while keeping their attributes up to date. Overall accuracy does not reveal how these demands interact. We introduce Audio
Memory Operations (AMemOps), a
benchmark for latest-attribute
retrieval, preservation of unchanged attributes and absent-source rejection in long audio streams. Paired controls vary recordings, training identities and earlier events to examine their effects on updating. Training on other recordings of the test instruments increases a reference delta-rule
memory's update advantage over matched
attention by 13.4 percentage points. At
retrieval delays of 120\,s, matched histories separate benefits of consistent repetition from costs of conflicting earlier values, while additive-
memory controls identify benefits of erasure.
Read the full abstract ▾
All authors Peiwei Chang, Li Zhang, Mark D. Plumbley, Philip J.B. Jackson, Wenwu Wang
Machine Learning
Does Pretraining Teach Chemical Language Models Chemistry?
Leif Sieben and Yoel Zimmermann
Chemical
language models (CLMs) are pretrained on large corpora of molecules with the aim of learning representations that transfer to diverse downstream tasks. Here, we systematically study the effects of pretraining using a series of such CLMs across 64 molecular property prediction
datasets. We find that the downstream gains of pretraining in CLMs do not scale and performance saturates below gradient-boosting baselines. Across the regimes evaluated, only supervised pretraining on traditional physico-chemical descriptors consistently improves downstream performance for the CLMs tested whereas some forms of pretraining result in models worse than a randomly initialized encoder. While models can resolve molecular structure well, they do so no better than a randomly initialized model, suggesting that, at best, CLMs learn descriptor-like properties in addition to low-level
token statistics during pretraining, but surprisingly little chemistry beyond what classical molecular features already capture.
Read the full abstract ▾
Machine Learning
Tiny Recursive Message Passing
Tommaso Salvatori, Matteo Forasassi, Alexis Toumi et al. (+1 more)
Message passing neural networks have shown remarkable performance in algorithmic and constraint
reasoning, due to their ability to encode the structure of the task directly in the architecture. Recent work has shown that better results can be achieved using recursive
transformers trained with deep supervision and truncated recursions, which make extremely deep loops stable to train, and equipped with test-time procedures such as discrete deduction
lattices. In this work, we apply the same procedures to message passing networks, and show that the two ingredients are complementary, and lead to parameter efficient models: we obtain about $99\%$ test accuracy on Sudoku-Extreme with a model of $16$k parameters that trains in about $10$ minutes on an H100, or in $9$ hours on a CPU. We further show that the cost of truncating gradients grows with the number of rounds information needs to cross the graph, making the choice of wiring a trade-off between robustness to truncation and transfer across problem sizes. We conclude by showing that our models transfer zero-shot to larger instances than those seen during training, outperforming popular neural baselines on N-queens and Boolean satisfiability with a fraction of their parameters.
Read the full abstract ▾
All authors Tommaso Salvatori, Matteo Forasassi, Alexis Toumi, Giovanni de Felice
Machine Learning
MultiAlloc: Budget-Aware Precision Allocation for KV Cache Quantization
Yangchengyu Zhou, Zhiping Wu, Runchang Jiang et al. (+5 more)
The Key-Value (KV) cache constitutes a dominant source of
memory consumption during large
language model (LLM) inference, particularly in long-context and complex
reasoning scenarios. Existing KV cache quantization methods have explored different compression granularities, including layer-wise precision adjustment and
token-level information preservation. In practical deployment, however, the key challenge is not merely compression, but how to allocate limited quantization resources among heterogeneous precision preservation strategies under a fixed KV cache
memory budget. We propose MultiAlloc, a profiling-guided precision allocation framework for KV cache quantization that adaptively distributes limited quantization resources between layer-wise precision enhancement and
token-wise preservation. MultiAlloc employs lightweight offline profiling to characterize layer sensitivity and
token importance, enabling adaptive precision allocation under a fixed
memory constraint. Guided by the profiled
memory–accuracy trade-off, it automatically derives an effective precision allocation strategy without online allocation search, and applies the resulting policy during inference. Extensive experiments on
reasoning and long-context
benchmarks demonstrate that MultiAlloc consistently improves over layer-wise,
token-wise, and recent multi-granularity KV cache quantization methods under matched
memory budgets across diverse models, including Mistral, Llama, DeepSeek-Qwen, and QwQ.
Read the full abstract ▾
All authors Yangchengyu Zhou, Zhiping Wu, Runchang Jiang, Dongdong Ren, Youfang Han, Wei Zhao, Dongqi Cai, Wenbin Li
Machine Learning · Cross-listed from cs.AI
To Align or Not To Align: Check Your COMPASS Before You Train
Fabian Gröger, Petar Damjanović and Maria Brbić
Combining pretrained models is a common strategy for building AI systems, whether by aligning a
vision encoder with a
language model or merging fine-tuned checkpoints. However, finding compatible pairs typically requires costly empirical search, as compatibility is known only after
alignment or merging. We propose COMPASS (COMPatibility ASsessment Score), a theoretically motivated diagnostic that estimates whether a model pair will be compatible prior to
alignment or merging. Building on the observation that representational convergence is most robust at the level of local neighborhoods, COMPASS decomposes model compatibility into two cheap-to-compute axes: a pair-specific ease, defined as the $k$-NN overlap between the two encoders, and a paradigm-level ceiling, capturing the maximum overlap their training paradigms allow. Together, these two axes provide a ranking of candidate model pairs and prescribe how to act on each pair: (i) prioritize pairs that already agree and whose paradigms are compatible, (ii) invest additional compute on pairs with compatible paradigms but low initial agreement, (iii) switch the pretraining paradigm when it cannot support sufficient agreement, or (iv) deprioritize pairs that score poorly on both axes. We apply COMPASS to $72$ cross-modal
alignment experiments and $392$ task-arithmetic model merges and show that compatibility estimates by COMPASS strongly correlate with downstream success, achieving average Pearson correlation of $0.91$ on cross-modal
alignment tasks and $0.70$ on model merging.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.RO
Grid or Nodes? Probing Object Memory in World Models Under Partial Observability
Justin McMillen and Yasin Yilmaz
A
robot's world model must remember objects that leave the camera's view. FloWM, a learned world model, keeps them in a spatial grid of features that it moves with five built-in velocities. A node
memory instead keeps each object's position and velocity as numbers. We compare them in FloWM's own block world, with an oracle-fed Kalman tracker as the node
memory. How the grid is read changes its score: a single-
cell probe finds 66% of out-of-view blocks, while FloWM's own probe 89%. FloWM also scores 8 points higher in view although no frame is shown, which we attribute to its write-back of its own predictions. Out of view, the tracker reaches FloWM's range in 14KB against 5.6MB, with no probe. On diagonal motion, outside its five velocities, FloWM finds 11% of blocks even after fine-tuning, while the unchanged tracker keeps 93%. Faster blocks defeat both until one more allowed speed takes the tracker to 94%, where FloWM's built-in shifts would need 25 copies of its grid. A learned node
memory must still replace the oracles and handle velocity under noise, long gaps, and the bounce.
Read the full abstract ▾
Machine Learning
Geometric Calibration of Supervised Fine-Tuning
Jianzhu Bao, Jia-Chen Gu, Qingyuan Liu et al. (+4 more)
Recent studies investigate the weight geometry through its spectral structure, which characterizes how the weights transform different input directions. Within this view, supervised fine-tuning (SFT) is shown to induce dense updates that perturb the spectral geometry of pre-SFT weights along their principal directions, potentially disrupting the model's prior knowledge. These observations motivate using the geometry of pre-SFT weights as a reference for refining learned SFT updates. We study this perspective as
Geometric Calibration and develop SPACE Subspace Projection for Adapted Checkpoint Enhancement), a lightweight post-hoc method that selectively reverts SFT update components within the principal subspace of the pre-SFT weights while preserving those in its orthogonal complement. SPACE is a data-free, training-free method that requires only the pre- and post-SFT checkpoints. We evaluate SPACE on Qwen2.5, Qwen3, Qwen3.5, Llama-3.1, and Olmo-3 models ranging from 2B to 14B, across mathematical, coding, and medical SFT, as well as several SFT variants. Overall, SPACE mitigates out-of-domain (OOD) capability regression while further improving in-domain (ID) performance over the uncalibrated SFT checkpoints. For math SFT, SPACE yields average ID/OOD gains of 2.85/2.26 points on Qwen2.5-7B, and 2.41/1.52 points on Qwen3-14B-Base, respectively. These results highlight how the
geometric relationship between SFT updates and pre-SFT weights shapes the adapted model's capabilities and, more broadly, motivate geometry-aware parameter updates for continual model improvement.
Read the full abstract ▾
All authors Jianzhu Bao, Jia-Chen Gu, Qingyuan Liu, Haozhen Zhang, Kai-Wei Chang, Nanyun Peng, Wenya Wang
Machine Learning · Cross-listed from cs.CV
MajorTOM:Elliot Pretrain, A Multi-Sensor, Multi-Temporal Pretraining Dataset for Earth Observation
Carmen Oliver Huidobro, Mikolaj Czerkawski, Kevin Monsalvez-Pozo et al. (+3 more)
Pretraining generalist Earth observation foundation models requires data that combine several sensors, varied temporal sampling and, increasingly, text, yet open
datasets rarely provide all three in one consistent, easily accessible form. We introduce MajorTOM:Elliot Pretrain, a collection of 279,166 tiles of 10 km drawn from the Major-TOM grid by clustering AlphaEarth
embeddings. Each tile stacks Sentinel-2 L1C, Landsat-8/9, Sentinel-1 RTC, Copernicus DEM and ESA WorldCover on a shared 1056x1056-pixel, 10 m grid, complemented by cloud masks, OpenStreetMap entities and ERA5 weather. The tiles are partitioned into three temporal regimes, Monotemporal (250,000 tiles, one acquisition), Monthly (12,500 tiles, twelve acquisitions spread over the annual cycle) and Burst (16,666 tiles, six acquisitions within 30 days), and are released as 6.3 TB of Cloud Optimized GeoTIFFs indexed by TACO v3 Parquet metadata, a standard that allows cloud streaming, efficient data access, and filtering before any pixel is read. A companion library derives sensor-tagged fact sheets, captions and 16 families of
vision-
language tasks from these layers at load time. We test our design choices in short pretraining runs of OlmoEarth-Base. First, a 25% subset of Major-TOM Core-S2L2A curated with AlphaEarth
embeddings matches or exceeds the full corpus on 8 of 9
benchmarks. Second, on Elliot-Pretrain, adding Landsat-8/9 improves or matches Sentinel-2 only pretraining on every
benchmark. Third, using DEM and WorldCover as prediction targets during pretraining further improves 8 of 9
benchmarks, while giving the same maps to the encoder as inputs degrades 5 of 9. Fourth, Sentinel-1 helps on 7 of 9
benchmarks, particularly on
benchmarks with radar inputs. Fifth, rule-based captions match a small
language model at no cost. The
dataset is openly available at https://source.coop/major-tom/elliot-pretrain.
Read the full abstract ▾
All authors Carmen Oliver Huidobro, Mikolaj Czerkawski, Kevin Monsalvez-Pozo, Cesar Aybar, Gustau Camps-Valls, Oscar J. Pellicer-Valero
Machine Learning · Cross-listed from math.ST
Minimax rates for pointwise CATE estimation: the role of design smoothness
Yuxuan Hou and Zhenghan Li
We study the role of design smoothness in the minimax estimation of conditional average treatment effects at a fixed
covariate point. For design densities bounded above and away from zero, without further regularity, we derive upper and lower risk bounds with an arbitrarily small gap in the convergence-rate exponent. Under uniform design, we obtain matching upper and lower bounds at interior points in the regime of low nuisance smoothness. These results show that rough
covariate designs can lead to slower rates than uniform designs, even under the same smoothness assumptions on the treatment effect and nuisance functions, highlighting the role of the
covariate distribution in pointwise causal estimation.
Read the full abstract ▾
Machine Learning
Deterministic Prediction as Global Velocity: Flow Matching for Probabilistic Time-Series Forecast
Jiajing Hu, Xu Han and Li-Ping Liu
Accurate uncertainty quantification is critical for many time series forecasting applications, and recent studies have increasingly adopted generative models to capture predictive uncertainty. Existing approaches often combine deterministic forecasting with generative modeling to improve predictive performance. However, their joint training is typically unstable because the deterministic model is embedded within the generative process, resulting in deep computational graphs and changing the probability transport path. In this work, we propose DV-Flow, a unified flow-matching formulation that incorporates the deterministic predictor as a parallel velocity component rather than as the endpoint of the transport path. This design enables stable end-to-end optimization while subsuming several existing formulations as special cases, allowing direct comparisons among them. Extensive experiments on six real-world
datasets demonstrate that DV-Flow consistently improves probabilistic forecasting performance over strong baselines. Our code is available at https://github.com/tufts-ml/dv-flow-time-series.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.AI
Can an Agent Learn to Predict Its Own Failure? Outcome-Grounded Confidence Calibration in LLM Agents
Ashutosh Hathidara
Controlling an acting LLM
agent means deciding, at each step of a long-horizon task, whether to trust its next action, abstain, or hand off to a human. That decision needs a per-step confidence signal predicting whether the current rollout will eventually succeed. The
agent's own self-reported verbalized confidence is the natural candidate, but existing reinforcement learning methods calibrate it against the correctness of a decision at the current step. We propose a method to calibrate it against the eventual outcome of a multi-step episode: under GRPO, the group of rollouts for a task supplies success labels, and a proper scoring rule scores each step's confidence against the realized outcome. In offline RL on recorded trajectories of a Qwen2.5-7B
agent in ALFWorld, TextWorld, and WebShop environments, the AUROC of step confidence against episode outcome rises from near chance to 0.68-0.85, and the top-10% of steps the
agent is most confident about succeed ~4.3x more often than average. In online RL on the live ALFWorld
agent, four calibration rewards improve early-step AUROC, expected calibration error goes from 0.42 to below 0.12, and selective execution improves over both the untrained
agent and a format-only RL control, without degrading task success. We also introduce an evaluation protocol that separates episode depth, task difficulty, and attempt-level foresight, which shows where the gain lies: the trained
agent recognizes from its first step which tasks it will fail, while telling a succeeding attempt from a failing one at the same task is harder, with some early-window signal emerging on TextWorld.
Read the full abstract ▾
Machine Learning
Jason Ken Adhinarta, Raina Wu, Ashutosh Sharma et al. (+5 more)
Self-
distillation is an emerging approach that trains a model to imitate its own predictions when conditioned on privileged information, such as correct answers or environmental feedback. Recent methods have reported better sample efficiency and generalization than reinforcement learning (RL), but it remains unclear whether these gains extend beyond in-distribution evaluations and whether they stem from privileged
distillation. In this paper, we revisit two representative methods, Self-
Distillation Policy Optimization (SDPO) and On-Policy Self-
Distillation (OPSD), across code, science, tool use, and math
reasoning tasks. We find that their reported gains do not consistently generalize to unseen problems and are largely explained by overfitting or transferring existing
reasoning behavior, rather than effectively using privileged information. Moreover, self-
distillation gradients show weak
alignment with policy gradients that directly optimize expected reward, questioning whether they provide an effective direction for policy improvement. Our experiments across Qwen3 models (1.7B-14B) and OLMo-3-7B suggest that effectively leveraging privileged information to improve
reasoning remains an open challenge. Our code is available at https://github.com/sleepypandu/does-sd-improve-
reasoning
Read the full abstract ▾
All authors Jason Ken Adhinarta, Raina Wu, Ashutosh Sharma, Avinash Balakrishnan, Rameswar Panda, Polina Golland, Yoon Kim, Zhang-Wei Hong
Machine Learning · Cross-listed from cs.SD
PhysicsSliders: Learning Physical Controls for Audio Foley Generation
Peiwei Chang, Yongqiang Chen, Yuelan Cheng et al. (+6 more)
Physical properties provide intuitive controls for sounds made by interacting objects. The challenge is to make a generative model respond predictably to these controls, both individually and in combination. We present \method\ (PXS), trained on simulated contact sounds with 15 material, geometry and contact parameters. Two separately trained variants accept direct parameter inputs or numerical captions. PXS controls hardness, size, speed and damping, supports room descriptions and produces spectrally distinct outputs at fixed targets. When controlling damping and size together, specifying thickness improves frequency accuracy and joint target delivery. Measurements from real recordings define effective-damping targets, which simulator inversion maps to generation inputs without real-audio fine-tuning. Across 18 impact classes from five real-audio
datasets, PXS-Physical follows these targets with mean grouped Spearman correlation $\rho=0.976$ and class-median error 0.094 dex (a factor of 1.24).
Read the full abstract ▾
All authors Peiwei Chang, Yongqiang Chen, Yuelan Cheng, Liting Gao, Yishun Lu, Li Zhang, Mark D. Plumbley, Philip J.B. Jackson, Wenwu Wang
Machine Learning · Cross-listed from cs.RO
ACCRUE: Continual Off-Policy Reinforcement Learning for Robot Foundation Models
Changyeon Kim, Kimin Lee, Yashraj Narang et al. (+1 more)
Reinforcement learning (RL) post-training improves
robot foundation models (RFMs) beyond their demonstrations, but real-world deployment demands more: learning new tasks efficiently while retaining previously acquired behaviors. Existing continual RL methods on top of VLAs mitigate forgetting but rely on on-policy updates with limited forward transfer, whereas off-policy methods improve single-task sample efficiency without addressing learner persistence across tasks. We propose ACCRUE, a continual off-policy RL framework that carries lightweight actor and critic networks across tasks, allowing each stage to reuse behavioral and value knowledge acquired through prior RL. The inherited critic adapts only on current-task transitions and guides current-task actor learning, while prior behavior is preserved by distilling the actor toward cached targets on a small number of retained episodes. Fixed visual features keep the inherited learner and its replay compatible with base-policy updates. Across contact-rich assembly sequences in simulation and on real-
robot hardware, ACCRUE accelerates subsequent-task learning with near-zero forgetting and significantly reduces real-
robot training time by 31.0% to 68.5%.
Read the full abstract ▾
All authors Changyeon Kim, Kimin Lee, Yashraj Narang, Yijie Guo
Machine Learning · Cross-listed from cs.CR
Autonomous Cyber Operations Gym for Training and Evaluating AI Models for Defence
Tobias Hürten, Johannes F. Loevenich, Florian Spelter et al. (+5 more)
This paper proposes an Autonomous Cyber Operations (ACO) gym, a unified framework integrating high-fidelity network emulation, realistic user behaviour modelling for benign network traffic and
adversarial profiles, and the generation of cybersecurity scenarios through open toolkits. By supporting live and deferred feature extraction at the packet and flow levels, the gym facilitates the creation of labelled synthetic
datasets for training Deep Neural Network (DNN)-based Intrusion Detection Systems (IDSs), as well as developing and evaluating modular components for autonomous defence
agents through emulated interactions. While evaluation of a DNN classifier within the gym shows high in-domain performance, achieving near-perfect recall on
adversarial network flows, our experiments reveal a significant sensitivity to domain shift, showing that even minor changes in network composition, such as variations in traffic protocol distribution, can lead to a catastrophic collapse in model transferability and detection performance.
Read the full abstract ▾
All authors Tobias Hürten, Johannes F. Loevenich, Florian Spelter, Erik Adler, Laurin Holz, Maxime Schwarzer, Thies Moehlenhof, Roberto Rigolin F. Lopes
Machine Learning
Fairness under Prediction Error in Predict-then-Optimize Pipelines
Hilda Adwubi Osei
Fairness constraints in predict-then-optimize pipelines are necessarily enforced on predicted outcomes $\hat{y}$, since true outcomes $y$ are unobserved at decision time. We study whether a demographic parity constraint enforced this way continues to hold against $y$, and how the resulting violation scales with prediction error. Using a budget-constrained allocation problem solved as a linear program, we show empirically that the true-outcome
fairness gap grows with prediction error $\sigma$ and saturates, well fit by $c(1 - e^{-\sigma/\tau})$ with $c = 0.1545 \pm 0.0103$ and $\tau = 3.9045 \pm 0.3745$; its mean is invariant to sample size while its variance decays at the standard $O(1/\sqrt{n})$ rate. We derive a two-part theoretical bound: a closed-form local sensitivity result showing the gap's first-order response to prediction error depends on only two individuals per group, validated against the solver to within three significant figures; and a universal, error-independent ceiling following from the boundedness of the decision variable, confirmed consistent with but looser than the empirical saturation level. On US Census income data, enforcing demographic parity on a trained predictor's output reduces an unconstrained disparity in true income to roughly one-third of its original size, leaving 34.4% $\pm$ 3.3% (California) and 35.7% $\pm$ 1.8% (Texas) of the disparity intact and undetectable from the predictions alone – consistent not only in mechanism but in approximate magnitude across two states with different demographics and income distributions, and stable across ten independent train/test splits in each. A
fairness constraint that appears fully satisfied at decision time can mask a substantial, systematic violation of the outcome it was meant to protect.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.SD
GeoMER: A Joint Euclidean-Hyperbolic Geometry Framework for Multimodal Emotion Recognition
Ting Zhu, Huizhi Liang, Boguslaw Obara et al. (+2 more)
Multimodal Emotion Recognition (MER) aims to understand human emotions by integrating information from multiple modalities such as
speech and text. Psychological studies suggest that emotions exhibit latent hierarchical structured relationships in their semantic representations. However, most MER approaches rely on Euclidean representations, which capture fine-grained cross-modal interactions but do not explicitly provide a
geometric prior for organizing hierarchical emotional structures. To mitigate this gap, we propose GeoMER, a joint Euclidean-
Hyperbolic modeling framework that introduces
hyperbolic emotion-core routing to capture higher-level, abstract emotional representations for MER. The proposed framework jointly models
token-level cross-modal semantic interactions and latent emotion representations in two complementary
geometric spaces. Specifically, GeoMER consists of a Euclidean
Multimodal Fusion Module (EuclFusion) for
multimodal representation fusion and a
Hyperbolic Geometric Core Routing (HypGeoRoute) module that performs geometry-aware core assignment and aggregation in a
hyperbolic manifold to capture more abstract feature representations. The HypGeoRoute module complements the Euclidean
token-level modeling. We experiment on the IEMOCAP
multimodal MER
benchmark and demonstrate that the proposed method achieves the best performance, with 83.54% weighted accuracy and 83.35% unweighted accuracy, outperforming Euclidean baselines, including MemoCMT.
Read the full abstract ▾
All authors Ting Zhu, Huizhi Liang, Boguslaw Obara, Shufei Duan, Varun Ojha
Machine Learning
When More Objectives Make Identification Easier: The Role of Preferences and Feedback
Bo Xue, Ji Cheng, Yuanyu Wan et al. (+1 more)
Best-arm identification (BAI) aims to identify the optimal arm using as few samples as possible. While most existing work focuses on a single objective, many real-world applications involve multiple criteria, which motivates the study of BAI in multi-objective bandits. Intuitively, multi-objective identification seems harder, since the learner must identify optimal arms in a higher-dimensional reward space; however, the additional objectives also provide richer feedback that can reduce the required exploration. In this paper, we characterize when and why multiple objectives can make BAI easier. For Gaussian bandits, we establish a characteristic-time lower bound for fixed-confidence identification that holds uniformly across preference orders. Through explicit comparisons under Pareto, linear, and lexicographic preferences, we reveal three phenomena: (i) identifying a larger optimal set can require fewer samples than identifying a unique optimum; (ii) the advantage of vector feedback over coordinate feedback depends fundamentally on the preference order; and (iii) multi-objective identification can require fewer samples than even the easiest corresponding single-objective problem. Finally, we propose Order-Aware Track-and-Stop (OATS), an
algorithm that is correct at any prescribed confidence level and asymptotically matches the lower bound.
Read the full abstract ▾
All authors Bo Xue, Ji Cheng, Yuanyu Wan, Shuang Qiu
Machine Learning · Cross-listed from cs.AI
LLM Agents for Automated Post-Training via Structured Recipe Optimization
Ruohong Liu, Bohan Tang, Lekang Jiang et al. (+2 more)
Automated post-training transforms pre-trained large
language models (LLMs) into models capable of handling diverse domains by automatically constructing and optimizing a recipe of coupled decisions over data sources, data processing, cross-domain composition, and training configuration under a limited experimental budget. Existing automation either searches predefined recipe spaces that enable systematic optimization but remain fixed rather than adapting their available choices to the target tasks, or uses open-ended
agents that discover task-specific choices but leave them implicit in trial-level edits, limiting their systematic reuse and recombination across trials. To fill this gap, we introduce optimization over a task-adapted recipe space, a novel paradigm that can adapt the recipe space itself to the target tasks while representing its choices explicitly for systematic reuse, recombination, and joint optimization. We instantiate this paradigm with Structured Recipe Optimization (SRO), comprising an LLM-based Recipe-Space Designer that constructs the task-adapted recipe space and a Multi-Domain Recipe Optimizer that uses structured trial evidence to guide LLM candidate construction and multi-objective
Bayesian trial selection. Extensive experiments show that SRO achieves the highest average score in all evaluated target-model and
agent-LLM combinations. In joint post-training across three capability domains and six
benchmarks, SRO reaches 41.1 versus 15.4 for the strongest baseline on Qwen3-1.7B with GPT-5.4, and 53.8 versus 40.8 on Qwen3-4B with DeepSeek-V4-Flash. Our code is available at https://github.com/RuohLiuq/SRO.
Read the full abstract ▾
All authors Ruohong Liu, Bohan Tang, Lekang Jiang, Jack Umenberger, Yiwen Guo
Machine Learning · Cross-listed from cs.GT
Linear Last-Iterate Convergence Rates for Optimistic Learning in Harmonic Games
John Lazarsfeld, Anas Barakat, Jacob Abernethy et al. (+4 more)
We prove that the last iterate of Optimistic Follow-The-Regularized-Leader (OFTRL) converges at a linear rate to a Nash
equilibrium in multiplayer harmonic games. Our result holds under a mild game-regularity condition and for a large family of regularizers, including
entropy, Tsallis, and log-barrier. Our analysis is based on a dual perspective of OFTRL: we prove that the interaction between optimism and the global structure of harmonic games leads to the dissipation of an energy function along the dual iterates, for which we establish tight quantitative bounds. Our result also holds for the Extragradient FTRL variant.
Read the full abstract ▾
All authors John Lazarsfeld, Anas Barakat, Jacob Abernethy, Vidya Muthukumar, Georgios Piliouras, Andre Wibisono, Juba Ziani
Machine Learning
Last-Iterate Convergence of Clipped SGD with Linear Decay
Wei Jiang, Haomin Bai and Zechao Li
In this paper, we study the last-iterate convergence of clipped
stochastic gradient descent under heavy-tailed noise. We assume that the
stochastic gradients are unbiased and that the noise has a bounded central $p$-th moment, where $1<p\le2$. For smooth convex objectives, a linearly decreasing learning rate gives a high-probability error bound of order $O(LR^2\log(1/\delta)/T+\sigma R(\log(1/\delta)/T)^{(p-1)/p})$. We further establish last-iterate guarantees for nonsmooth composite objectives and smooth composite objectives with proximal updates. In addition, when a causal adversary replaces at most $q$ gradients, the same clipped update achieves a bound with $\log(1/\delta)$ replaced by $\log(1/\delta)+q$. The corresponding lower bounds for the clean and corrupted oracle models are also provided.
Read the full abstract ▾
Machine Learning
Unbalanced W-Flow: Unbalanced Wasserstein Drifting under Finite-Batch Marginal Fluctuations
Jongwook Jeon, Mingue Park, Jiwon Choi et al. (+3 more)
Teacher-free one-step generative models such as W-Flow learn directly from distributional objectives, with the Sinkhorn divergence providing the strongest empirical performance among the energies studied in W-Flow. However, Sinkhorn divergence is based on balanced optimal transport, whose exact marginal constraints can induce undesirable transport under finite minibatch fluctuations. We introduce UW-Flow, which replaces the balanced Sinkhorn energy with a debiased unbalanced Sinkhorn divergence and derives the corresponding Wasserstein velocity. The resulting dynamics retain W-Flow's cross-minus-self structure, recover the balanced limit, and preserve target identifiability under suitable assumptions. We further characterize the finite-batch velocity error of balanced and unbalanced transport by comparing each minibatch field with its population counterpart. Our analysis identifies conditions under which marginal relaxation yields smaller minibatch-to-population velocity error, and shows that for separated matched modes, balanced sensitivity grows with inter-mode separation while the corresponding Unbalanced Optimal Transport (UOT) sensitivity remains controlled by local transport geometry. Controlled experiments on fixed-support and continuous two-mode distributions support these predictions. On MNIST, CIFAR-10, and ImageNet-100, UW-Flow improves generation under matched training budgets, with the largest gains in unconditional generation and consistent improvements in the class-conditional setting.
Read the full abstract ▾
All authors Jongwook Jeon, Mingue Park, Jiwon Choi, Jaihoon Kim, Yunhong Min, Minhyuk Sung
Machine Learning · Cross-listed from stat.ML
Understanding Circulant Permutation to Extend the CMinHash Estimator
Keegan Kang, Avery Hood and Benedict H.W. Wong
CMinHash promises a new direction for designing MinHash
algorithms with reduced variance and storage space via circulant permutations. However, the original variance analysis is difficult to extend to other MinHash
estimators. We present a new approach that gives explicit and easily comparable approximate variance expressions, and use it to analyze the Minner
estimator. Numerical experiments show that our variance expression matches the empirical mean square error (MSE) of estimates. Our work thereby demonstrates how to apply circulant permutation to other MinHash
estimators and compute their approximate variance.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.CV
Beyond the Encoder: Object-Centric Representations Emerging from Segmentation Foundation Models
Michalis Pistos and Bartłomiej W. Papież
Foundation models are increasingly used as feature extractors for downstream tasks with limited data, typically through representations obtained from their encoders. However, foundation models trained for structured tasks such as segmentation also form object-specific representations during task execution, whose transferability remains largely unexplored. We investigate Segment Anything Model 2 (SAM2) object pointers, compact object-level representations produced through the segmentation pathway, as transferable representations for downstream tumour prediction. Through experiments on two PET/CT tumour prediction tasks, we find that object pointers learned using segmentation supervision alone already encode tumour-discriminative information, despite never being optimised for the downstream labels. We further introduce task-guided object pointers, where an auxiliary objective jointly shapes the object pointer space for the downstream prediction together with segmentation. Task-guided pointers improve cancer-type classification, outperforming conventional encoder-based transfer and radiomics, while segmentation-only pointers perform best on the more challenging recurrence prediction task. Representation analyses reveal associations between object pointers and interpretable radiomics features, class-dependent organisation of the learned representation space, and explicit spatial grounding around the segmented tumour. Together, these findings suggest that useful transferable representations can emerge through the execution of a foundation model's primary task, providing an alternative to conventional encoder-based transfer.
Read the full abstract ▾
Machine Learning
Euclidean Distance-Based Flow Matching
Paul-Henrik Heilmann and Peter Ochs
Flow Matching has recently become popular as a technique for generative modeling with applications in many domains. In the standard Flow Matching formulation, we interpret the vector field as the solution of a least-squares regression problem, which is known to be sensitive to contamination. Therefore we leverage the viewpoint of robust statistics and replace the squared Euclidean distance with the Euclidean distance as a regression objective. This seemingly minor change strongly affects the properties of Flow Matching. We provide a probabilistic interpretation of the new regression target as a conditional
geometric median and use this characterization to study distributional robustness of the Euclidean Distance-based Flow Matching. We analyze the mathematical properties of this changed objective and the effect of a contaminated target distribution on the involved conditional velocity distributions. We discuss the challenge of directly characterizing the distribution that our method induces. In specific contamination settings we show theoretically that our method yields an improvement for severe contamination. Finally, we present numerical experiments conducted on the CIFAR-10 image
dataset where a proportion of the samples has been contaminated. Our Euclidean distance-based Flow Matching shows a clear advantage in scenarios of severe contamination of the target distribution.
Read the full abstract ▾
Machine Learning
Two For The Price Of One: Curvature-Aware Pruning Preserves Influence Functions Too
Mehdi Makni, Natalia Ponomareva and Rahul Mazumder
Post-training
pruning enables efficient deployment, but conventional methods do not explicitly preserve internal-gradient
alignment between sparse and dense models. We study novel connections between downstream performance of compressed models and the ability to retain gradient information via influence-function preservation. Central to this connection is a
curvature-aware objective that accounts for both layer activations and backpropagated gradients. Despite its promise, this objective remains underexplored in
pruning due to a lack of efficient solvers. We introduce $\texttt{CACTUS}$, a novel ADMM-based optimizer supporting both
pruning and sparse-plus-low-rank decompositions of pre-trained weights. Compared to LLM-Surgeon which targets the same objective, $\texttt{CACTUS}$ can achieve over an order-of-magnitude speedup. At $50\%$
pruning of Qwen3-4B, $\texttt{CACTUS}$ reduces the C4 perplexity gap to the dense model by $45\%$, improves compressed-to-dense per-sample gradient cosine similarity on WikiText-2 from $0.670$ to $0.755$, and runs over $28\times$ faster than LLM-Surgeon on substantially weaker hardware. By preserving influence functions, $\texttt{CACTUS}$ supports effective data selection. On Llama-3.2-3B, fine-tuning a fresh dense model on the $5\%$ of data selected by $\texttt{CACTUS}$ achieves $49.44\%$ accuracy on BIG-Bench Hard, compared with $47.22\%$ using ALPS and $49.54\%$ using dense-model selection. Finally, integrating $\texttt{CACTUS}$ with LoGRA, a recent gradient-projection strategy, on Qwen3-4B with MLP layers pruned to 2:4 sparsity and a rank-64 correction retains a $0.711$ cosine
alignment with dense LoGRA-projected C4 gradients.
Read the full abstract ▾
Machine Learning
Approximating Optimal Policies in Heterogeneous Asynchronous Multi-player Bandits
Sacha Arroues-Paykin, Etienne Boursier and Vianney Perchet
Motivated by applications to cognitive
radio networks, multi-player multi-armed bandits (MMAB) have attracted considerable
attention in recent years. The asynchronous variant, in which players are active only at a fraction of timesteps, better reflects real-world decentralized networks, where devices are not permanently active. Despite its practical relevance, this setting remains relatively understudied. Existing work on asynchronous MMAB has focused on homogeneous players, who share the same activation probability, a setting in which the optimal centralized allocation can be easily characterized. In contrast, efficiently computing the optimal allocation in the heterogeneous setting, where players have different activation probabilities, has remained an open problem. In this work, we investigate this problem and first establish its NP-hardness. We then turn to the greedy
algorithm, which runs in
polynomial time, and prove that it achieves a $\frac{1}{e}$-approximation of the optimal allocation. Furthermore, when players are processed in decreasing order of their activation probabilities, the approximation guarantee improves to $\frac{1}{\sqrt{e}}$. Overall, our results establish the greedy
algorithm as an efficient approximation to the optimal allocation, providing a practical oracle baseline for the design of online
algorithms for heterogeneous asynchronous multi-player bandits.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.CR
Towards Safe Autonomy for Computer Use
Dominik Seip, Bartol Markovinović and Matthias Hein
The safety of LLM-based
agents currently rests on
alignment training and safety classifiers, neither of which can rule out that a sophisticated indirect
prompt injection (IPI) goes through and causes significant harm. Recent work therefore builds
agents with information-flow guarantees that rely on structured tool interfaces to interact with the environment. Computer Use
Agents (CUAs) act visually on the computer screen, which does not provide a boundary between trusted first-party interfaces and untrusted, potentially attacker-authored content. Actions are equally ambiguous: a click may land on an application toolbar or a button controlled by an attacker. Treating the user as the source of trust and confirming each sensitive decision produces confirmation fatigue, with the risk of missing harmful actions. We introduce SA-CUA, a dual-LLM framework that generates a reliable trust boundary for CUAs by deriving conservative deterministic rules from the accessibility tree for unmasking trusted regions from the privileged LLM's screenshot, keeping trusted and untrusted content separated. Human interaction is only required for proper task specification in the beginning. During the task a robust endorsement channel allows values to pass from the untrusted to the trusted domain by verifying information via trusted channels. This allows the
agent to autonomously handle tasks involving untrusted information. SA-CUA maintains much higher utility than CaMeL-CUA while keeping the attack success rate at zero.
Read the full abstract ▾
Machine Learning · Cross-listed from stat.ML
How Low Can You Go: Fine-Tuning LLMs with Bayesian Optimization (LlMBO)
Aditya Kulkarni, Davide Sartor, Carlos Soto et al. (+1 more)
Recently, Morris et al. introduced TinyLoRA, a strategy for LLM fine-tuning which introduces arbitrarily few parameters, and demonstrated accuracy improvements using less than $10$ parameters rather than the ${\gtrsim 10,000}$ typically demanded by LoRA-based methods. However, that study continued to use gradient-based local search methods of the kind deployed for training multi-billion parameter LLMs in the first place. This article is predicated on the observation that, in the low dimensional setting, global optimization methods become tractable, unlocking Large
language Model
Bayesian Optimization (LlMBO). LlMBO is sample-efficient, gradient-free,
memory-light, parallelizes across devices without synchronization, and, unlike policy-gradient methods, can optimize greedy-decoding accuracy directly. We describe numerical experiments comparing LlMBO against GRPO on fine-tuning tasks for a variety of models on a math
benchmark suite which show that LlMBO can be a suitable alternative in this regime, matching or exceeding GRPO's fine tuning gains, while using a fraction of the compute and
memory.
Read the full abstract ▾
All authors Aditya Kulkarni, Davide Sartor, Carlos Soto, Nathan Wycoff