Machine Learning
Train for Many, Update with One: Improving Multi-Sample Reasoning via Fenchel-Bregman Posterior Learning
Sungjun Lim, Adrian Robert Minut, Nico Carlsen Daheim et al. (+3 more)
Multi-sample
reasoning measures candidate coverage, the probability that at least one of $K$ candidates is correct. However, minimizing the average loss of individual attempts does not directly optimize candidate coverage. We therefore formulate finite-$K$ posterior learning, which trains a distribution over solver parameters to reduce the probability that all $K$ independently generated candidates fail. To optimize this objective efficiently, we propose a Fenchel-Bregman (FB) approach that requires only one posterior draw per update, rather than the $K$ used by direct Monte Carlo optimization. FB first uses an exact Fenchel reformulation to separate problem-specific weights from the posterior update, enabling one-draw posterior updates. It then applies a closed-form Bregman update to track these weights through an exponential running average of observed failures, reusing the same draw. We instantiate FB with IVON, a variational optimizer for Gaussian posteriors over solver parameters, yielding Fenchel-Bregman IVON (FBI). Across diverse recursive and autoregressive
reasoning tasks, FBI improves candidate coverage. In a matched comparison, it achieves comparable Pass@10 using 12.3% of the direct Monte Carlo baseline's measured training time per update.
Read the full abstract ▾
All authors Sungjun Lim, Adrian Robert Minut, Nico Carlsen Daheim, Mohammad Emtiyaz Khan, Kyungwoo Song, Thomas Möllenhoff
Machine Learning
Explore with Memory, Learn without It: In-Context Exploration and Self-Distillation for LLM Reasoning
Yutong Chen, Zhichao Xu, Rui Liu et al. (+1 more)
Exploration bounds learning efficiency in reinforcement learning with verifiable rewards (RLVR). Current methods explore mostly through independent sampling, without fully using the information gathered along the way, such as known wrong answers and environment feedback. We propose LEAF, which injects a dynamic, problem-specific
memory into the context before every rollout. The
memory contains the most recent response, its reward, and other environment feedback. It is updated after every verifier call. We decouple policy updates from rollouts and update the policy through self-
distillation. On four
benchmarks and two open models, LEAF accelerates exploration by about 4x compared with a random sampling baseline. On unseen test sets, it improves accuracy over the strongest baseline by 1.59 and 1.83 percentage points on the two models, while using less compute than any RL baseline. These results highlight the potential of LEAF in improving both exploration and learning efficiency. Code is available at: https://github.com/on1262/LEAF
Read the full abstract ▾
All authors Yutong Chen, Zhichao Xu, Rui Liu, Kunpeng Liu
Machine Learning
CoCA: Criterion-Aware Credit Assignment for Rubric-Based Reinforcement Learning
Guangcheng Zhu, Yingfan Ma, Xing Zheng et al. (+7 more)
Rubric-based reinforcement learning (RL) evaluates responses against explicit criteria to construct rewards, offering a viable way to train large
language models on open-ended tasks. However, standard methods compress criterion-level feedback into scalar rewards and assign uniform
token advantages, leaving structured rubric supervision underutilized. Recent work seeks finer-grained guidance by treating rubrics as privileged information to derive self-
distillation targets, but the resulting learning signal is not grounded in criterion outcomes and can lead to reward degradation. This motivates us to reframe the problem as criterion-aware credit assignment, where privileged rubric information serves as a relevance cue for credit allocation rather than a source of imitation targets. Accordingly, we propose CoCA, which derives criterion-level credits from group-relative outcomes and allocates them across
tokens based on criterion-induced distribution shifts, yielding both outcome-grounded and fine-grained learning signals. Empirically, CoCA achieves an average gain of +4.2% over standard rubric-based RL across medicine and science
benchmarks, demonstrating more effective learning on open-ended tasks.
Read the full abstract ▾
All authors Guangcheng Zhu, Yingfan Ma, Xing Zheng, Haobo Wang, Shenzhi Yang, Kai Tang, Zhengqing Zang, Bowen Song, Weiqiang Wang, Gang Chen
Machine Learning · Cross-listed from cs.CV
DT-LoRA: Reaching Beyond the Tangent Space of Low-Rank Adaptation
Yangming Zhang, Ling Yi, Datong Li et al. (+4 more)
Low-rank adaptation (LoRA) often underperforms full fine-tuning. We investigate this performance gap through a tangent-space lens: factorized updates are confined to the tangent space of the current low-rank factors, leaving gradient components outside this space unexploited. An analysis of ten representative LoRA variants reveals that these residual components account for most of the gradient energy and exhibit strong spectral concentration. Motivated by this finding, we propose Dynamic Tangent LoRA (DT-LoRA), which partitions a fixed rank budget into a stable branch and a dynamic branch to capture the principal residual gradient directions. We show that these directions define the optimal update of a local quadratic model under orthogonal-complement and rank constraints.
Matrix-free randomized SVD recovers them without materializing the residual gradient, while controlled write-back preserves accumulated updates and stabilizes training within the allocated rank. DT-LoRA adds no trainable parameters or inference cost over LoRA. Experiments across computer
vision and natural
language processing tasks demonstrate consistent improvements over LoRA and its variants. Moreover, despite being a general-purpose adaptation method, DT-LoRA outperforms task-specific approaches to open-vocabulary domain generalization for semantic segmentation, achieving state-of-the-art results. Code is available at https://github.com/Snipine/DT-LoRA.
Read the full abstract ▾
All authors Yangming Zhang, Ling Yi, Datong Li, Jinliang Ding, Xiaojie Wang, Tiecheng Song, Zhaolong Ning
Machine Learning
TSRL: Trigger-based Supervised Reinforcement Learning for Continual Agent Training
Yuqian Fu, Kaiwen Jiang, Zijun Song et al. (+3 more)
Large
language model (LLM)
agents excel at complex, multi-step tasks but suffer from catastrophic forgetting during continual learning–acquiring new skills often degrades previously mastered capabilities. Existing post-training paradigms face a fundamental trade-off: Supervised Fine-Tuning (SFT) is sample-efficient but prone to severe forgetting, while Reinforcement Learning (RL) preserves prior capabilities but converges slowly without expert guidance. To bridge this gap, we introduce Trigger-based Supervised Reinforcement Learning (TSRL), a framework that integrates the strengths of both paradigms through an adaptive, state-dependent expert supervision mechanism. TSRL employs RL as the optimization backbone, activating SFT selectively on "hard states" where the policy fails. Theoretically, we show that restricting supervision to failure states derives an upper bound on the cumulative parameter updates, thereby reducing forgetting compared to full SFT. Empirically, TSRL achieves consistent gains across three
agent benchmarks (ALFWorld, WebShop, and Deep Search) at both 7B and 32B scales. In continual learning scenarios, our method retains up to 9.6 points more capability on prior tasks than SFT->RL baselines, while matching or exceeding target-task performance.
Read the full abstract ▾
All authors Yuqian Fu, Kaiwen Jiang, Zijun Song, Jiawei Wang, Yuanheng Zhu, Dongbin Zhao
Machine Learning · Cross-listed from cs.AI
Evidence-First Agentic Verification
Reza Sanayei, Razvan-Gabriel Dumitru, Sunjie Hou et al. (+7 more)
Long-horizon, semi-verifiable agentic tasks often rely on rubric-based evaluation, but the evidence needed to assess these tasks can be spread across the environment state and the execution trace. A judge may overlook this evidence or accept unsupported claims, including an
agent's own reports of success. When rubric verdicts serve as reinforcement-learning rewards, such errors can reinforce undesirable behavior. We introduce Sauron, an evidence-first agentic verifier that investigates both state and trace and relies on a claim only after independently validating it against its source. Bounded queries and joint investigation of related criteria reduce context use and redundant work, and each verdict carries an auditable receipt of its supporting evidence. Across 2,586 human-labeled criteria from 100 personal-assistant episodes, Sauron achieves the highest F1 of five verifiers on all eight backbones tested while costing less than Gandalf, the strongest baseline. With GPT-5.6 Sol, it reaches 0.96 F1 versus Gandalf's 0.92 at 23% lower cost, with a wider margin on process criteria. On a paired fault-injection suite covering 13 families, several of which simulate reward hacking, Sauron detects 97% of injected failures while flagging only 2% of clean controls, outperforming baselines even when they are given the full trace. On TRAIL, a
benchmark for detecting and localizing real
agent errors, Sauron leads both Gandalf and TRAIL's own judge on all three quality metrics.
Read the full abstract ▾
All authors Reza Sanayei, Razvan-Gabriel Dumitru, Sunjie Hou, MohammadHossein Rezaei, Haris Riaz, Robert Vacareanu, Minglai Yang, Ying Liu, Mihai Surdeanu, Daniel Yue Zhang
Machine Learning · Cross-listed from quant-ph
Representation Geometry and Inductive Biases in Quantum Classifiers across Classical Visual Benchmarks
Yu Yamaoka, Weng Ian Chan and Kosuke Mitarai
Quantum Machine Learning (QML) has a theoretical advantage over classical
algorithms for specific synthetic data that is computationally hard. However, it remains unclear whether such theoretical advantages play a meaningful role for realistic classical data, such as natural images.
Quantum structural properties induced by
unitary operations or
entanglement introduce inductive biases that differ from those of classical models, potentially leading to distinct learning behavior on real-world data. In a fully
quantum learning pipeline, classical real-world data are first encoded into
quantum states via a
quantum feature mapping, analogous to a classical feature extractor, and a
quantum classifier then classifies the resulting states.
Quantum feature mappings generate representations that are difficult to realize classically, while also inducing representation-level biases that do not necessarily match the structure of realistic data. To enable a fair comparison at the classifier level, we focus on the classifier and use classical feature extractors that already provide powerful representations for natural images. By controlling the data scale and supervision strength, we analyze how the
unitary structure of
quantum classifiers affects generalization, and show that their effectiveness depends strongly on the geometry of the input representations. These results indicate that
quantum classifiers benefit from classical features that are aligned with
quantum feature spaces. Our analysis shows that kernel
alignment identifies representations compatible with
quantum feature spaces, providing insights into translating classical models into
quantum algorithms.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.AI
ECG-RSI: Recursive Self-Improvement through Targeted Electrocardiogram Waveform Evolution
Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng et al. (+8 more)
Electrocardiogram (ECG) models need supervision that targets evolving weaknesses and remains faithful to waveform evidence. We introduce ECG-RSI, a recursive self-improvement framework combining verified waveform synthesis with feedback-guided intervention search. Signal remeasurement verifies waveform edits; construction errors guide generator repair, while learner errors motivate interventions compared on a separate selection set. Across seven
benchmarks, three rounds raise Qwen3.5-35B-A3B's mean question-answering accuracy from 34.57% to 56.04% and mean classification F1 from 24.27 to 34.03. Across the retained trajectory, the preferred supervision shifts from direct reinforcement to confusion-focused contrasts. Continued feedback correction yields diminishing external gains. These results motivate separating error diagnosis from intervention selection and checking whether gains on the feedback set transfer to independent
benchmarks.
Read the full abstract ▾
All authors Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Jiangwei Lao, Quanbin Wang, Honglin Lin, Jiajia Liu, Zhihong Lu, Hao Jin, Jian Wang, Lijun Wu
Machine Learning
MissPath-FM: Flow Matching with Structured Priors for Partially Observed Time Series
Genpei Zhang
Flow matching has emerged as a powerful alternative to
diffusion models for generative modeling, but standard conditional flow matching is fundamentally misaligned with partially observed time series: starting from an uninformed Gaussian source forces the model to transport even already-observed entries from noise to the target, wasting capacity on trivially known regions. We propose MissPath-FM, a partial-observation-aware flow matching framework that redesigns the source prior and probability path for incomplete time series. MissPath-FM employs structured source priors (interpolation priors and Gaussian process posterior priors) to initialize the flow close to the observed signal, paired with a missingness-aware smooth path that allocates transport asymmetrically between observed and missing regions. We evaluate systematically across four
datasets, four missingness mechanisms (MCAR, block, MNAR, irregular), two missing ratios, and three sequence lengths (T in {24, 48, 96}), totaling 48 settings against eight baselines. MissPath-FM wins 29 of 32 settings at T=48 and 45 of 48 overall (93.8%). Our ablations reveal a regime-dependent design hierarchy in which no single component suffices: under random missingness the source prior yields a large initial gain, but under structured block missingness it can fail entirely and the learned flow becomes indispensable; the missingness-aware path contributes a further 12%-26% across all regimes. Mechanistically, structured priors shorten observed-region transport by up to 15x, concentrating the velocity field on genuinely uncertain entries. Project page: https://zgp110.github.io/misspath-fm/
Read the full abstract ▾
Machine Learning
Confine to the Task, Anchor the Rest: Concept-Guided Retention Regularizer for Parameter-Efficient Fine-Tuning
Ibna Kowsar, Ivan Medri, Fanyu Meng et al. (+1 more)
Parameter-efficient fine-tuning (PEFT) has been widely used, yet the adapted model suffers catastrophic forgetting of pre-trained knowledge. While most existing work hardens the adapter by constraining where the weight update may lie, we note that constraining this weight channel does not constrain the activation channel. In this paper, we propose Concept-Guided Retention (CGR), a plug-and-play regularizer that constrains both for any PEFT base. CGR reads a frozen sparse autoencoder (SAE) and treats two disjoint concept sets differently. It softly confines the update to the task-concept subspace (via weight channel), and anchors a selected set of general-purpose concepts to the frozen model (via activation channel). We also provide anchor value bounds based on weight updates to understand how they interact, and show that both terms are necessary. On a CLIP
vision panel, CGR significantly improves retention at comparable accuracy for every base, and the same pattern holds for world-knowledge retention under Llama-3.1-8B instruction tuning. Because the protected directions are named SAE concepts, CGR remains inspectable and controllable.
Read the full abstract ▾
All authors Ibna Kowsar, Ivan Medri, Fanyu Meng, Xin Liu
Machine Learning
Where Small Mistakes Matter: Quantifying State Criticality for Imitation Learning
Yusuf Umut Ciftci, Joonwon Kang, Ajay Suresha Sathya et al. (+1 more)
Imitation learning typically treats every demonstration state as equally important, optimizing a behavior-cloning objective uniformly across the
dataset. Yet not all states are created equal: in a small number of critical states, even a slight deviation from the expert action can lead to a failure, while in many other states such errors are recoverable and inconsequential. We formalize state criticality in imitation learning as the worst-case degradation in downstream value induced by a bounded perturbation of the expert action at a given state. We then propose SCRIM, a practical State CRIticality Measure that estimates this quantity by training an
adversarial disturbance policy and a value critic around a fixed imitation policy. SCRIM assigns each demonstration state a score reflecting how vulnerable the task outcome is to small, policy-relevant action errors. We show that these scores can be used to reweight behavior-cloning losses, improving imitation at consequential states and increasing overall task performance. We further show that SCRIM enables data curation by selecting compact, high-value subsets of demonstrations that preserve performance while substantially reducing the amount of training data. We evaluate SCRIM across various simulated and real-world robotic manipulation tasks, and demonstrate that SCRIM improves policy training, enables efficient data selection, and provides interpretable insights into which moments in a demonstration are most critical for task success. Project website: https://sia-lab-git.github.io/SCRIM/
Read the full abstract ▾
All authors Yusuf Umut Ciftci, Joonwon Kang, Ajay Suresha Sathya, Somil Bansal
Machine Learning
Focus on What You Need: Global-to-Local Dynamic Attention for Multimodal Sarcasm Detection
Dehao Wu and Xingjie Zhuang
Multimodal sarcasm detection has attracted increasing
attention due to the widespread use of images and text on social media platforms. Existing studies generally treat the semantic or emotional inconsistency between images and text as a key cue for identifying sarcastic expressions. However, images in real-world scenarios often contain multiple visual objects, many of which are irrelevant to sarcasm-related semantics. Such redundant visual information may introduce noise during cross-modal interaction, distracting models from focusing on critical sarcasm cues and potentially leading to inaccurate predictions. To address this issue, we propose GLDA (Global-to-Local Dynamic
Attention), a
multimodal sarcasm detection framework that models cross-modal interactions from a global-to-local perspective. Specifically, GLDA first employs a global
multimodal fusion module to capture high-level semantic interactions between visual and textual modalities, enabling the model to identify coarse-grained cross-modal inconsistencies. To further reduce the influence of irrelevant objects, GLDA introduces a local gated cross-modal
attention mechanism that dynamically filters redundant visual and textual features and refines
multimodal representations at a finer granularity. In addition, a multi-branch learning objective is designed to jointly optimize unimodal and
multimodal representations. Experimental results demonstrate that GLDA effectively improves sarcasm detection performance and achieves competitive results compared with existing approaches. The code will be available at https://anonymous.4open.science/r/GLDA-DBE6.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.CV
A Differentiable 3D Scene Graph Metric via Contextual Hellinger Triplet Geometry
U-Chae Jun, Yoonhyuk Choi and Jiwoo Kang
3D scene graph prediction is commonly supervised with local object and predicate classification losses. Although effective for slot-wise label prediction, this formulation does not provide a graph-level semantic distance between a predicted scene graph and its ground truth: semantically mild label errors and structurally disruptive relational errors can be treated similarly, and subject-predicate-object facts are not compared as scene-level units. We propose Contextual Hellinger Triplet Geometry, a graph-level supervision objective that measures the semantic and structural discrepancy between 3D scene graphs. Our key idea is to reinterpret the conventional classifier outputs of a 3D scene graph predictor as an aligned probabilistic scene graph field, where object and predicate predictions jointly define contextual subject-predicate-object facts. This formulation yields a differentiable metric-based training objective with an efficient factorized implementation and can be applied to existing 3D scene graph prediction architectures without modifying their network design. Experiments on 3DSSG demonstrate that our supervision improves downstream graph-conditioned 3D scene generation, while ablation studies confirm the complementary roles of the proposed graph-level terms and user studies show better graph descriptiveness, relation plausibility, and structural preservation.
Read the full abstract ▾
Machine Learning · Cross-listed from cs.AI
Understanding Gene Expression-to-Image Learning: Spatial Structure and Information Fidelity
Xiaoqing Lian, Pei Liu, Jiahui Tan et al. (+3 more)
Single-
cell RNA sequencing (scRNA-seq) enables
cell-level analysis of biological heterogeneity, but learning effective representations from high-dimensional, unordered gene expression profiles remains challenging. Gene Expression-to-Image (GE2I) offers an effective learning strategy by arranging gene values on pixel grids for convolutional neural networks, but its underlying mechanisms remain unclear. We study GE2I on full-transcriptome data while keeping the network and training procedure fixed. Our results reveal a trade-off between spatial smoothness and gene-level fidelity. Structured mappings produce smooth, low-frequency patterns that support faster early optimization, whereas assigning multiple genes to the same pixel obscures expression differences and can reduce predictive performance. This trade-off explains why random mappings can compete with structured layouts when many genes are represented. One-to-one gene-to-pixel assignment improves prediction across the tested visual backbones even when image smoothness decreases. We further find that the benefits of pretraining and augmentation depend on the backbone and mapping. These findings provide practical guidance for pixel allocation and the transfer of
vision priors to gene expression data.
Read the full abstract ▾
All authors Xiaoqing Lian, Pei Liu, Jiahui Tan, Quan Zou, Chen Lin, Xiangxiang Zeng
Machine Learning
LeanPRM: Lean-Grounded Process Supervision for Natural Language Proof Verification
Rodney Lafuente-Mercado, Salman Rahman, Jingyan Shen et al. (+3 more)
AI systems now write long and difficult mathematical proofs, and checking their step-by-step validity is becoming a bottleneck. A learned verifier could check proofs cheaply, but training one requires a correctness label on every step, and such labels are hard to obtain for open-ended proofs. Existing methods take them from human annotators, which is costly, or derive them from final answers, which proofs do not have, and formalizing each proof in Lean requires expert effort. We introduce LeanPRM, a framework that uses the Lean proof assistant to create step-level labels for natural-
language proofs. A model writes a natural-
language proof together with a paired formal proof, without feedback from Lean. Lean then labels every step, and a second model discards failures caused only by the formal syntax, so that the negative labels mark errors in the mathematics. LeanPRM contains 10,634 proofs with 152K labeled steps and no human annotation. A Qwen3-4B verifier trained on LeanPRM raises exact step-label accuracy on a held-out set of 200 proofs (LeanPRM-Bench) from 48.2 to 74.7, as opposed to 60.3 when the same base model is trained on PRM800K. It also selects better proofs in best-of-N selection on IMO-level problems, and outperforms the PRM800K-trained verifier on three of four external verification
benchmarks. Training on LeanPRM improves verifiers at every size from 0.6B to 8B parameters on LeanPRM-Bench, Hard2Verify and IMO-GradingBench. Given the labels need no human annotation, LeanPRM can grow with the supply of formal statements, offering a path to step-level feedback on AI-generated proofs that are hard for humans to check.
Read the full abstract ▾
All authors Rodney Lafuente-Mercado, Salman Rahman, Jingyan Shen, James Shiffer, Saadia Gabriel, Pavel Izmailov
Machine Learning
zPocket: Better Pocket Finding with Frozen Co-Folding Features
Tally Portnoi, Allison M Keys, Benjamin R DiFrancesco et al. (+1 more)
Identifying
protein-ligand binding sites (pockets) is important for understanding
protein function and developing drugs. However, no public method has leveraged co-folding models for novel pocket prediction; furthermore, this task lacks a suitable
benchmark as existing
benchmarks are within the training distribution of co-folding models. We therefore developed PocketWatch,a curated
dataset of biologically relevant ligand pockets. The PocketWatch test set consists of unseen pockets stratified by sequence similarity, structural similarity, and crypticity to enable evaluation of generalization to novel and challenging pockets. Using this
dataset, we develop zPocket, a method that trains pocket-prediction heads on top of frozen features from co-folding models. We show that zPocket heads outperform baseline methods across all stratifications. Our best zPocket model achieves overall per-target AUPRC of 0.745 compared to 0.598 and 0.595 for the strongest baselines, VN-EGNN and UniSite-1D retrained on the PocketWatch split (0.559 and 0.531 when using the published weights); the widely used P2Rank reaches 0.477. We further show that zPocket can be used to outperform baseline methods on top-N+2 recall of predicted pocket centroids through clustering per-residue predictions or using zPocket to rescore centroids proposed by other models. Finally, we explore two applications of zPocket: crystallographic fragment screens and a proteome-scale scan for uncharacterized pockets on human
proteins. Together these results suggest that zPocket heads built on co-folding trunk features can greatly advance ligand binding site prediction.
Read the full abstract ▾
All authors Tally Portnoi, Allison M Keys, Benjamin R DiFrancesco, Nate Gruver
Machine Learning · Cross-listed from cs.AI
MedEvo: Medical Reasoning Evolution through Knowledge Navigation
Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng et al. (+10 more)
Medical
multimodal reasoning models are typically trained on teacher-generated rationales, discarding the valid
reasoning in students' own failed trajectories, while general-domain correction methods offer little guidance for errors in image reading, anatomical context, or differential diagnosis. We propose MedEvo, which uses medical knowledge, namely a tag taxonomy and error-type diagnosis, to navigate both trajectory repair and data selection. For each failed rollout, a gold-blind auditor locates the first consequential error and provides a minimal, answer-free hint; the same student retains the valid prefix and regenerates the erroneous suffix. After each round, MedEvo identifies capability gaps across medical tags on a held-out feedback set and prioritizes related questions for new rollouts and repairs. With supervised fine-tuning (SFT) alone, MedEvo raises Qwen3.5-9B mean accuracy across seven
benchmarks from 65.57% to 69.12% (47.40% to 54.50% on the
multimodal track of MedXpertQA (MedXpertQA-MM)), surpassing teacher-
distillation SFT (66.93%). It also raises Qwen3.5-35B and Qwen3.5-397B mean accuracy from 67.91% to 73.04% and from 71.83% to 74.95%, respectively. Our analyses show that failed trajectories contain
reasoning worth preserving, that repair success does not imply learning value, and that independent feedback is key to sustained improvement.
Read the full abstract ▾
All authors Xiaoliang Wang, Zhuo Chen, Chuanyang Zheng, Quanbin Wang, Jiangwei Lao, Yingao Liu, Jiajia Liu, Jinming Cao, Honglin Lin, Zhihong Lu, Hao Jin, Jian Wang, Lijun Wu
Machine Learning · Cross-listed from cs.CL
Encoding Context into LLM Weights: What Makes It Context?
Yue Wang, Qizhou Wang, Feng Liu et al. (+3 more)
Longer contexts allow large
language models (LLMs) to condition their predictions on more information, but inference costs grow prohibitively with context length. An emerging alternative uses test-time updates to encode context as temporary changes to model weights. However, encoding this information into weights that already represent general knowledge obscures its role as context, so the LLM may fail to use it properly to condition subsequent inference. To restore this contextual role, we first associate each context with an identifier during encoding, thereby distinguishing the newly encoded information from pre-existing knowledge. Invoking the identifier at inference then enables the LLM to recite the encoded context. Nevertheless, successful recitation does not ensure that the LLM will use the recalled information as context. We therefore further introduce a prompting framework that explicitly guides the LLM to use this information as context. Across diverse controlled tasks and practical
benchmarks, our method outperformed the baseline using only off-the-shelf LLMs, showing its advances in context encoding.
Read the full abstract ▾
All authors Yue Wang, Qizhou Wang, Feng Liu, Gang Niu, Masashi Sugiyama, Bo Han
Machine Learning · Cross-listed from cs.CL
LatentRiskBench: Measuring State-Dependent Risks in Enterprise Agents
Jingyue Cong, Estrid He, Zhuang Li et al. (+5 more)
Enterprise
agents face safety risks that cannot be determined from individual actions alone. Existing safety
benchmarks largely evaluate action-local risks, overlooking failures whose safety depends on workflow state. In such state-dependent risks, the same action may be safe in one context but violate policy in another. Hence, we introduce LatentRiskBench for evaluating such state-dependent risks in enterprise settings. It pairs risky trajectories with benign counterparts that share the same terminal action but differ in latent policy-relevant states, requiring models to reconstruct workflow context rather than rely on action-level cues. We also introduce new metrics for state-dependent risk evaluation, including a suite of residual risk (RR) measures. Experiments on the
benchmark show that action-level evaluation detects only 0.09 of state-dependent violations, whereas state-level evaluation reaches 0.96. Existing guardrail methods largely overlook such risks unless they are explicitly applied to workflow states. Case studies on a commercial deployed scenario further show that the safety metrics embedded in the
benchmark provide a more realistic measure of safety for enterprise
agents. Our
dataset and code are available at https://github.com/JingyueCong/LatentRiskBench.
Read the full abstract ▾
All authors Jingyue Cong, Estrid He, Zhuang Li, Nan Wu, Li Shen, Yameng Peng, Xinshu Li, Andy Song
Machine Learning · Cross-listed from eess.IV
Deficit-Driven Stratified Acquisition for Building STRATA-S12: An Open SAR-Optical Corpus for Self-Supervised Learning
Christian Sanchez-Leon, Dongming Chen and Dongqi Wang
Building a co-registered SAR-optical corpus for self-supervised Earth Observation is above all an acquisition problem, yet the main multi-modal
datasets are collected opportunistically, with no per-stratum quota control and no mechanism against the geographic dropout that non-uniform cloud cover leaves behind. We present Deficit-Driven Stratified Acquisition (DDSA), an
algorithm that treats building a co-registered Sentinel-1/Sentinel-2 corpus as a constrained maximum-coverage problem over a hemisphere-corrected (class x season x region) stratification, enforces strict temporal pairing and a multi-stage quality gate, and closes the loop during acquisition with a human-in-the-loop deficit-repair stage. Measured on its own product, the
algorithm yields three results a single-pass acquisition cannot report. Deficit-driven ordering makes a partial run a covering one: it holds every one of the 192 strata populated from the first few percent of the budget, where an abundance-first order leaves most strata empty until the end. Whether a failed attempt can be recovered depends on the temporal persistence of its cause, not the kind of defect:
geometric failures such as nodata are among the most repairable (eta = 41.4%) while cloud is among the least (eta = 18.8%), because a cloudy location is cloudy on most dates of its season. The repair loop supplies 39.5% of the released pairs at no measurable quality cost, with 89.8% surviving manual review. The quality guarantee has a price: 2,928,370 acquisition attempts produced 740,746 released pairs. We release the corpus, STRATA-S12, with STAC/PDS4-aligned FAIR sidecars, and a pre-registered linear-probe check confirms it is fit for self-supervised pre-training. A metadata-level coverage comparison against SSL4EO-S12 and SEN12MS situates the corpus over a shared frame; systematic cross-corpus and cross-method comparison is left to follow-up work.
Read the full abstract ▾
Machine Learning
Dynamic K-Head: Difficulty-Supervised Budget Selection for Continuous Latent Reasoning
Ying Li, Yuqi Chen, Sixuan Li et al. (+1 more)
Chain-of-thought (CoT) improves
reasoning but incurs substantial autoregressive rationale cost. Continuous CoT moves intermediate computation into recurrent hidden states, yet typically uses a fixed number of updates even though the best budget varies across questions and accuracy need not improve monotonically with $K$. We introduce Dynamic K-Head, a question-conditioned, pre-rollout budget selector. It is jointly trained with the reasoner under selected budgets, using difficulty supervision based on scores assigned from the question alone. We also show that latent-update count can misrepresent end-to-end cost because a short latent rollout can shift computation into a longer explicit continuation. We therefore count latent updates, implementation-specific special
tokens, and all decoded text. Instantiated with CoDI on Llama 3.2 1B and GSM8K, Dynamic K-Head reaches 54.44% exact-match accuracy at 12.03 average
reasoning/output
tokens, compared with 52.31% at 15.21
tokens for fixed-$K=6$ CoDI—a 2.12-point gain with 20.9% lower
token cost. Across three out-of-distribution arithmetic
datasets, it reduces cost by 19.3–25.7% while improving accuracy by 0.30–2.22 points. These results demonstrate that difficulty-supervised budget selection can improve both accuracy and efficiency, while stable budget selection remains a central challenge.
Read the full abstract ▾
All authors Ying Li, Yuqi Chen, Sixuan Li, Yunfeng Cai
Machine Learning
DeltaTTG: Training Multi-Turn Agents with a Token-to-Go Progress Meter
Xuyan Ye, Xiyan Fu, Shengda Fan et al. (+3 more)
Reinforcement learning with verifiable rewards (RLVR) for multi-turn
agents relies on sparse terminal rewards that say little about individual steps. Within-group step comparisons depend on reliable state matching and sufficiently diverse sampled outcomes, while direct step scoring may require costly external evaluators, human annotations, or task-specific heuristics. We introduce DeltaTTG to learn a reusable progress signal through a shared progress meter, using low-cost supervision from diverse policies' recorded trajectories without task-specific progress rules. Its supervision target,
Token-to-Go (TTG), the number of
tokens remaining to be generated until task completion, serves as an observable proxy for remaining completion effort. During actor training, changes in the frozen meter's estimates redistribute outcome-derived advantages across steps without altering the reward, preserving each trajectory's advantage sign and
token-average scale. On ALFWorld, PaperSearchQA, and BFCLv3, DeltaTTG improves the average score over GRPO by 5.7 and 7.2 points with Qwen3-4B and Llama-3.1-8B-Instruct, respectively, outperforms all compared RL baselines on average, and keeps actor-training GPU-hours close to GRPO. Further experiments show improved sample efficiency and out-of-domain performance, and support the meter's ability to distinguish step quality and guide test-time action selection across actor sizes. The code is available at https://github.com/Xuyan923r/DeltaTTG.
Read the full abstract ▾
All authors Xuyan Ye, Xiyan Fu, Shengda Fan, Jiawei Cui, Yankai Lin, Wei Lu
Machine Learning · Cross-listed from stat.ML
textGrain watermarking is harder to detect: can sum-tests still do it optimally?
Swapnaneel Bhattacharyya and Soham Bonnerjee
Watermarking embeds an imperceptible signal in text from large
language models that is statistically detectable with a secret key. The recent textGrain watermark (Li et. al. (2026b)) calibrates this signal against the model's sampling
entropy through an explicit
entropy budget, combining the stochasticity of Red-Green and the unbiasedness of Gumbel-max watermarks, and its accompanying report proposes a detector that sums pivotal statistics. We study when, and how efficiently, textGrain can be detected after human edits, modeled as a sparse mixture in which a fraction $\varepsilon_n\asymp n^{-p}$ of positions retain the watermark and
token uncertainty decays as $n^{-q}$. First,
entropy calibration comes at a cost: the minimax detection boundary of textGrain is $p+q=1/2$, so its detectable region is strictly smaller than that of Gumbel-max and maximal-coupling watermarks, whose boundary is $2p+q=1$. This cost stems from textGrain's construction rather than from
entropy calibration itself: an unbiased scheme with the same
entropy loss remains detectable where textGrain is not. Second, perhaps surprisingly, this loss makes the sum detector optimal at the boundary when the number of transport columns is fixed, unlike for Gumbel-max. When the number of columns grows and the
entropy budget does not vanish, the detectable region expands while that of the sum detector does not. Third, for a fixed edit fraction, we characterize the optimal exponential rate of type II error decay and show that the sum detector falls strictly short of it at every fixed alternative. Finally, we propose a weighted multiscale binomial scan, valid in finite samples and needing no knowledge of the next-
token distributions,
entropy budget, or edit fraction; it attains every detection boundary above and the optimal exponent.
Read the full abstract ▾
Machine Learning
DisaggKD: Efficient Knowledge Distillation through Teacher-Student Disaggregation and a Fused Loss
Ruixuan Tu and Yuyang Dong
Knowledge
distillation (KD) frameworks place teacher and student on the same GPUs, so per-GPU
memory limits the teachers that fit. DisaggKD separates them across nodes: teacher replicas stream hidden states to the student, a KD loss over row windows and vocabulary shards equals the full-vocabulary loss in exact arithmetic without forming the full-vocabulary logits, and a prefetch ring paces producer and trainer. Closed-form
memory and traffic models locate the
memory limits of each layout. On the Qwen3.5-35B-A3B to 9B pair, the disaggregated layout matches the LM loss of the colocated run with the same loss and is faster than that run; our fused loss also lets a colocated run train on as few nodes at faster speed, while the split is
memory efficient when the student spans several nodes. On the larger pairs, colocated runs with a fused loss still run OOM where the split trains, because teacher weights and student states fill the budget.
Read the full abstract ▾
Machine Learning
Mixed Precision Scaling Laws for Quantized Pretraining
Jinuk Kim, Rajarshi Saha, Muhammed Fatih Balin et al. (+1 more)
Since scaling laws were introduced for modern LLMs (Kaplan et al., 2020), several works have extended them to account for quantization, tracking the emerging low-precision data types supported on AI accelerators. We develop a more accurate scaling law for quantized pretraining with mixed precision across the forward and backward passes, grounded in optimization theory and validated empirically. Our formulation separates the effects of quantization on model capacity and on optimization efficiency, and explicitly models the persistent optimization error from forward and backward quantization identified through convergence analysis. To select the regimes for fitting and evaluating our law, we first decompose existing 4-bit training recipes into their forward and backward components, isolating those that most affect loss. Our Qwen3 experiments under selected 4-bit regimes, together with published Llama-style experiments, show that our law predicts loss with up to $11\times$ lower MSE than existing laws as model size and
token count scale. Finally, we extend our law to layer-wise mixed precision and identify the optimal quantized-pretraining recipe, i.e., model size, training
tokens, and per-layer forward/backward precision, under a hardware-aware compute budget. Unlike prior precision scaling laws, our law predicts that the optimal precision increases with the compute budget.
Read the full abstract ▾
All authors Jinuk Kim, Rajarshi Saha, Muhammed Fatih Balin, Youngsuk Park
Machine Learning
The Tokens Strike Back: Safety-Aware Denoising with Directional Embedding-based Revision
Amman Yusuf, Jiwoo Shin, Ruizhi Deng et al. (+3 more)
The Safety-Aware Denoiser (SAD) is a recent, training-free, safety-guidance framework for text
diffusion models. During inference, SAD steers text samples toward safe regions of the
token probability space. In this paper, we propose $\mathit{SADDER}$ (Safety-Aware Denoising with Directional
Embedding-based Revision), which improves the safety and quality trade-offs of SAD by exploiting a safety-guidance signal in the
embedding (or activation) space for each
token position, then updating the
token positions using the SAD update rule multiplied by the guidance signal. This multiplication acts as a gating mechanism:
tokens with a high unsafe signal are more likely to be modified during inference following the SAD update rule. However, such an update does not automatically preserve the conditional mean, i.e., the denoiser of a fixed target distribution. We therefore theoretically derive new position-wise target distributions for which the modified denoiser is exact. Furthermore, we derive conditions under which the position-wise target distributions match a global target distribution. We test our $\mathit{SADDER}$ method on popular Masked
Diffusion Language Models (MDLMs) and Uniform-State
Diffusion Models (USDMs) across unsafe continuation and jailbreak settings, and show that $\mathit{SADDER}$ achieves a better balance between safety, generation quality, and inference cost than existing methods.
Read the full abstract ▾
All authors Amman Yusuf, Jiwoo Shin, Ruizhi Deng, Heitor Rodrigues Guimarães, Nikita Dvornik, Mijung Park
Machine Learning
Calibrated Disambiguation for Partial Multi-label Learning
Zhuoming Li, Yuheng Jia, Mi Yu et al. (+1 more)
Partial multi-label learning (PML) aims to train a classifier on
datasets whose instances are over-annotated with not only relevant labels but also irrelevant labels, which is common when
datasets are collected from crowd-sourcing platform. Existing works primarily approach it from a curriculum learning perspective, leveraging the memorization effect to disambiguate noisy labels and produce robust predictions. However, these methods are based on non-adaptive weighting functions and lack theoretical guidance for optimal weighting. To overcome these issues, a calibrated disambiguation model named PML-CD is proposed. We first formulate the optimal weighting function for curriculum-based disambiguation, which is equivalent to the calibration of the model's predicted confidences, thus providing guidance for curriculum designing. To obtain the optimal weighting function from PML
dataset during the training, a transferable calibrator is designed, which takes the histogram of positive samples' confidences as input, and outputs the optimal curriculum weighting for training. Prototype
alignment regularization is also proposed to promote the model's performance. Experiments conducted on Pascal VOC, MS-COCO, NUS-WIDE and CUB have verified that our method outperforms existing state-of-the-art PML methods.
Read the full abstract ▾
All authors Zhuoming Li, Yuheng Jia, Mi Yu, Zicong Miao
Machine Learning
RELAY: Selective Self-Distillation with Adaptive Hand-off for Rubric-Based Reinforcement Learning
Kunwang Zheng, Xinyu Sun, Yu Cao et al. (+7 more)
Rubric-based reinforcement learning grades a response against many written criteria but collapses them into one reward, which discards the detail the rubric carries and gives no GRPO reward gradient when a group's responses tie. A frozen teacher that reads the rubric as privileged context provides
token-level supervision, but we observe an early plateau in the gains over GRPO. RELAY uses advantages and rewards to select positive-advantage and low-scoring zero-advantage rollouts, and monitors the
distillation loss to decide when to retire the teacher. Stage 1 combines selective
distillation from a frozen teacher with GRPO; a plateau detector then removes the teacher, and Stage 2 continues with GRPO alone. Calibrated on one model-domain pair and reused unchanged in three others, the plateau detector fires between steps 122 and 171. RELAY scores highest in-domain among the compared methods in all four settings, ahead of GRPO on RaR-Medicine by 4.62 points with Qwen3-1.7B and 6.02 with Qwen3-4B. On Qwen3-1.7B/Medicine, removing the teacher from a shared checkpoint outperforms keeping, weakening or refreshing it. Distilling selected rollouts covering about half the response
tokens also outperforms distilling all rollouts at a matched average coefficient.
Read the full abstract ▾
All authors Kunwang Zheng, Xinyu Sun, Yu Cao, Shuocheng Li, Lei Cheng, Ying Liu, Qilong Zhang, Jinlong Chen, Hua Yang, Qianglong Chen
Machine Learning
FedLoRD: Personalized Federated Learning based on Low-Rank Structures of Model Parameter Differences
Tsutahiro Fukuhara, Junya Hara, Hiroshi Higashi et al. (+1 more)
We propose FedLoRD, a personalized
federated learning (PFL) method based on the low-rank structure of the differences between the global and personalized model parameters. PFL is a branch of
federated learning that aims to learn models tailored to the local data of each client under biased local data distributions. In its local training, each client learns a personalized model by regularizing the difference between its model parameters and the global ones sent from the server. The regularization function needs to be designed with priors on the model parameter difference; however, the existing PFL methods rely on empirical choices (e.g., $\ell_1$- and $\ell_2$-norms), which do not take into account the actual characteristics of the difference. The resulting mismatch of priors can lead to suboptimal personalized models. To address this issue, we propose a PFL method whose regularizer is built on an observed characteristic of the model parameter differences: they are approximately low-rank at every layer. In our method, we first formulate a local training problem with nuclear norm regularization on the layer-wise differences. We further rewrite this problem via the factorized representation of the nuclear norm so that it can be efficiently solved by a plain
stochastic gradient descent method. Image classification experiments on three
benchmark datasets demonstrate that FedLoRD outperforms several existing PFL methods.
Read the full abstract ▾
All authors Tsutahiro Fukuhara, Junya Hara, Hiroshi Higashi, Yuichi Tanaka
Machine Learning · Cross-listed from cs.CV
Morphology-aware deep learning predicts colorectal cancer survival from whole-slide images
Hikmat Khan, Metin N. Gurcan, Anil Parwani et al. (+3 more)
Background: Five-year survival prediction from hematoxylin-and-eosin whole-slide images could support risk stratification, but the task is weakly supervised, and the tissue morphologies carrying prognostic information remain incompletely characterized. Methods: We conducted a retrospective secondary analysis using TCGA-COAD/READ (n=562) for morphology-specific analysis, Alliance/ CALGB 89803 (n=424) for model development, and SurGen (n=218) for external evaluation. We propose MorphContrast, a morphology-aware contrastive multiple-instance learning framework that uses tissue labels as structural supervision for five-year survival prediction. MorphContrast combines gated-
attention pooling with three auxiliary objectives: patch-level morphology labels define within-patient contrasts, outcome concordance or discordance defines cross-patient contrasts within matched morphologies, and diversity regularization discourages representational collapse. Morphology-matched patches from patients with discordant outcomes serve as hard negatives, encouraging prognostic discrimination within comparable histologic contexts. Results: Morphology-specific models show weak but distributed prognostic discrimination: adipose tissue achieves an area under the receiver operating characteristic curve (AUC) of 0.622, stroma achieves 0.604, and both perform comparably to tumor epithelium (0.601-0.605). On Alliance/CALGB 89803, the single model achieves an AUC of 0.642, while the post hoc ensemble reaches a pooled out-of-fold AUC of 0.70 and a risk-group hazard ratio of 3.79 (95% confidence interval: 2.55-5.63). On the external SurGen colon cancer cohort (n=218), the single model and ensemble achieve AUCs of 0.669 and 0.678, respectively. Conclusions: These findings support tissue morphology as structural supervision for outcome-aware representation learning in colorectal cancer prognosis.
Read the full abstract ▾
All authors Hikmat Khan, Metin N. Gurcan, Anil Parwani, Wendy L. Frankel, Wei Chen, Muhammad Khalid Khan Niazi