arXiv subjects

Masashi Sugiyama

Publications and source records attributed to Masashi Sugiyama.

At least 19 recordsLinked to original sources

arXiv2026

Practical Estimation of the Bayes Optimal Fairness-Accuracy Tradeoff with Soft Labels

There is a fundamental limit to the best prediction error of any model for a given data distribution. In classification, the Bayes error quantifies this limit and serves as both a benchmark for trained models and a criterion to detect overfitting. The state-of-the-art Bayes error estimation techniques for binary classification use only soft labels (positive class probabilities) and require no input features or auxiliary training. We extend such notions to constrained classification problems, such as fairness. Fair classification is an important case of constrained or multi-objective classification. It requires a model not only to be accurate but also to yield fair outcomes (equal or near-equal true positive rates) across different demographic groups (e.g., race and gender). This leads to an inherent fairness-accuracy tradeoff or a Pareto frontier that is used to compare fair classifiers against each other. The focus of our work is to estimate the fundamental limit for this fairness-accuracy tradeoff. We derive an explicit expression for the fair Bayes error rate (i.e., best achievable error rate subject to given fairness constraints), and construct consistent, instance-free estimators for it using only soft labels and group information. We provably bound the bias and the sample complexity of our estimators. Empirically, our method produces a robust, low-variance estimate of the optimal fairness-accuracy tradeoff curve, even when the soft labels are noisy and derived from a classification model on hard labels.

cs.LG↗

arXiv2026

DELTA: Robustly Training Label-Conditional Diffusion Models with Weak Annotations

While label-conditional diffusion models have recently exhibited remarkable generative capabilities, their success heavily relies on massive, cleanly labeled datasets. In practice, categorical supervision is rarely perfect: it is often corrupted by noise, clouded by ambiguity, or partially missing. Training directly on such weak annotations severely degrades generation quality, and existing robust methods offer only fragmented solutions that depend on scarce auxiliary priors. To overcome these limitations, we propose DELTA, a unified framework that robustly trains Diffusion modELs across all three weak annoTAtion types. By treating the unknown true label as a latent variable, DELTA optimizes a principled variational objective that jointly recovers the clean data distribution and infers the true label posteriors. To make this joint training computationally tractable, we further introduce a median-centered timestep sampling strategy that efficiently concentrates estimation where the diffusion process is most informative. Extensive experiments on synthetic and real-world benchmarks demonstrate that DELTA produces high-fidelity, class-consistent samples across diverse weak supervision scenarios, outperforming specialized baselines while demanding strictly less prior knowledge.

cs.LG↗

arXiv2026

Adaptive Localization of Knowledge Negation for Continual LLM Unlearning

With the growing deployment of large language models (LLMs) across diverse domains, concerns regarding their safety have grown substantially. LLM unlearning has emerged as a pivotal approach to removing harmful or unlawful content while maintaining utility. Despite increasing interest, the challenges of continual unlearning, which is common in real-world scenarios, remain underexplored. Successive unlearning tasks often lead to intensified utility degradation. To effectively unlearn targeted knowledge while preserving LLM utility, it is essential to minimize changes in model parameters by selectively updating those linked to the target knowledge, thereby ensuring other knowledge remains unaffected. Building on the task vector framework, we propose a new method named ALKN (Adaptive Localization of Knowledge Negation), which uses dynamic masking to sparsify training gradients and adaptively adjusts unlearning intensity based on inter-task relationships. Comprehensive experiments across three well-established LLM unlearning datasets demonstrate that our approach consistently outperforms baseline methods in both unlearning effectiveness and utility retention under continual unlearning settings.

cs.LG↗

arXiv2026

Dataset Mismatch Matters in Group Relative Policy Optimization for Reinforcement Learning from Verifiable Rewards

Recently, reinforcement learning from verifiable rewards (RLVR) is practically very important, where the leading approach is group relative policy optimization (GRPO) that typically faces a trade-off between rollout efficiency and downstream performance. To improve efficiency while preserving performance, existing selective rollout methods mainly exploit only source-side training dynamics to optimize the discrete distribution over source prompts. However, in practice, the source training dataset is rarely identically distributed to the target test dataset. When such a mismatch exists, due to legal or privacy constraints, it is usually not acceptable to share target prompts themselves with the RLVR trainer. Nevertheless, it is still possible for the target data owner to provide limited aggregated feedback from a small held-out target set without revealing target prompts or rollout traces. In this paper, we propose learning-to-reweight rollout (L2RR) that explicitly incorporates target-side performance feedback into GRPO. Specifically, L2RR learns a discrete sampling distribution over source prompts by optimizing a bi-level objective that maximizes accuracy on a held-out target dataset, while simultaneously penalizing rollouts that provide little incremental learning information. Empirical studies on six math reasoning benchmarks and three model scales show that L2RR improves target-domain performance and increases rollout efficiency.

cs.LG↗

arXiv2026

Asymptotically Exact Negative Guidance of Diffusion Models via Positive-Unlabeled Learning

Negative guidance of diffusion models generates samples that avoid an undesired condition, specified either by reference samples or by a conditional score alongside an unconditional model. Existing methods inject repulsive terms derived from local information at each sample, so the resulting sampler does not target an explicit density, and its bias cannot be systematically reduced by increasing computation. We instead present a negative-guidance method that samples from an explicit target density, which uniformly handles both reference-sample and score-based specifications and suppresses probability mass near the undesired condition with an avoidance strength that parametrizes the target density itself. Realizing this target requires estimating, at each diffusion timestep, how likely each sample is to belong to the undesired condition. We cast this estimation as a Positive-Unlabeled learning problem that admits lightweight online training along the diffusion trajectory, using condition samples as positives and diffusion particles as unlabeled data. Because the guided particles drift from the unconditional marginal, we further correct the residual mismatch via Sequential Monte Carlo, yielding a sampler that is asymptotically exact in the particle limit and that turns any existing negative-guidance method into its proposal kernel. Empirically, our method approaches the ground-truth target on Gaussian mixtures and Pareto-dominates existing methods on trade-offs between avoidance and generation quality, distributional bias, and diversity, including in an image-generation setting with a pretrained score network.

cs.LG↗

arXiv2026

Rethinking Learning from Label Proportions via Moment Matching

Learning from label proportions (LLP) is a weakly supervised setting where training data are grouped into bags with only bag-level label proportions observed, and the goal is to learn a classifier that predicts labels for individual instances. Recently, a large body of methods has emerged; however, most of them rely on the assumption that instance-label pairs are sampled i.i.d. and randomly assembled into bags, an assumption often violated in real-world scenarios. Our experimental results indicate that these methods perform suboptimally under non-random settings. In this work, we adopt a more realistic assumption: bags are sampled i.i.d., and instance labels are conditionally independent given the instances. Under this setting, we show that the label counts follow a Poisson multinomial distribution. Motivated by this observation, we propose LLP via moment matching (LLP-MM), a simple yet effective approach which leverages multi-order factorial moments as training objectives, encouraging the classifier predictions to match these moments and thereby more fully exploit the underlying statistical structure of the data. Extensive experiments on benchmark datasets under various bag construction strategies demonstrate the effectiveness of our approach while maintaining high computational efficiency. Our code is available at https://github.com/TianhaoMa5/LLP-MM.

cs.LG↗

arXiv2026

Understanding Negative Learning in Large Language Model Fine-tuning

Negative learning has never been more important than it is today, particularly in large language model (LLM) fine-tuning, where it underpins machine unlearning, preference- and reward-based optimization, and exploratory data generation. Yet its functional role remains a subject of ongoing debate, with existing work variously characterizing it as forgetting, regularization, or target shaping. However, analytical support focuses primarily on the single-update scale, while empirical results at the trajectory scale are typically obtained under negative-only training and may not extend to composite objectives in which it interacts with other components. To develop a deeper understanding, we introduce a gradient-based endpoint probing framework that can examine negative updates separately throughout training. Our analysis reveals that negative learning plays different roles across algorithms and training phases. For example, in off-policy DPO, it acts more like a regularizer early in training and shows evidence of target shaping only toward out-of-distribution endpoints and only in later phases. In on-policy PPO and GRPO, by contrast, negative updates offset endpoint-aligned positive learning without stable target shaping, consistent with a role in sustaining exploration. Guided by these findings, we develop a family of stage-aware methods that modulate the strength of negative updates throughout training, improving fine-tuning performance across tasks.

cs.LG↗

arXiv2026

Encoding Context into LLM Weights: What Makes It Context?

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.

cs.CL↗

arXiv2026

DenMark: Robust Semantic Watermarking for Diffusion Language Models

Semantic text watermarks encode signals in meaning rather than surface token choices, offering robustness to paraphrasing and other semantic-preserving edits. Existing semantic watermarking methods are primarily designed for autoregressive language models (ARLMs), where completed candidate units can be generated and scored before generation proceeds. This paradigm does not naturally extend to diffusion language models (DLMs), where semantic units remain incomplete during intermediate denoising steps and tokens may be updated in flexible orders. We propose DenMark, a semantic watermarking framework that injects key-dependent signals directly into the DLM denoising process. DenMark partitions the output into fixed token regions and uses temporary rollouts as semantic lookahead: conditional completions estimate the eventual semantics of an incomplete region, enabling DenMark to select local updates with higher estimated semantic watermark scores. Repeating this procedure across denoising steps progressively accumulates watermark evidence in the final output. For detection, DenMark uses calibrated scanning over candidate unit sizes to remain robust to boundary shifts introduced by semantic attacks. Across four DLMs, three datasets, and four semantic attacks, DenMark gives the strongest overall robustness and achieves the best TPR at each reported operating point in 46 of the 48 backbone-dataset-attack combinations. Overall, DenMark provides an effective semantic watermark for DLMs. Code is available at https://github.com/TianhaoMa5/DenoMark.

cs.CL↗

arXiv2026

Preemptive LLM Unlearning against Forbidden Capability Acquisition via Gradient Sealing

Open-weight LLMs are released not only as fixed products but also as substrates for downstream fine-tuning. This openness, however, creates legal and ethical risks because users may misuse fine-tuning to instill illicit knowledge or enable hostile operations. Model providers therefore need apre-release defense against such acquisition, motivating the problem of preemptive unlearning. Unlike retrospective unlearning, which removes capabilities already present in a fixed model, preemptive unlearning seeks to prevent their acquisition under unseen attack data and future fine-tuning procedures. Despite its practical importance, this setting remains largely unexplored, presents distinct challenges, and is therefore the central focus of our work. We first verify that existing retrospective methods provide insufficient pre-release protection. Even when forbidden capabilities are suppressed in current outputs, forbidden-domain data can still induce gradients through internal pathways, enabling later acquisition. Motivated by this finding, we propose a gradient-sealing principle that blocks these pathways by pushing relevant pre-activations into the negative region, where ReLU-family activations exhibit zero or near-zero derivatives. Experiments across multiple LLM families demonstrate our stronger resistance to downstream acquisition than retrospective baselines, validating gradient sealing as an effective mechanism for pre-release protection.

cs.LG↗

arXiv2026

Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers

Reinforcement Learning with Verifiable Rewards (RLVR) replaces costly human labeling with automated verifiers. To reduce verifier hacking, many RLVR systems binarize rewards to $\{0,1\}$, but imperfect verifiers inevitably introduce \emph{false negatives} (rejecting correct answers) and \emph{false positives} (accepting incorrect ones). We formalize verifier unreliability as a stochastic reward channel with asymmetric noise rates $ρ_0$ and $ρ_1$ -- the FP rate and the FN rate, respectively. From this abstraction we derive two lightweight corrections: (i) a \emph{backward} correction that yields an unbiased surrogate reward and thus an unbiased policy-gradient estimator in expectation, and (ii) a \emph{forward} correction that reweights score-function terms so the expected update aligns with the clean gradient direction and requires only the FN rate. We implement both as lightweight hooks in a group relative policy optimization pipeline, both corrections improve RLVR for math reasoning under synthetic and real verifier noise, with the forward variant being more stable under heavier noise. Finally, an appeals mechanism with a lightweight LLM verifier estimates the FN rate online and further improves performance.

cs.LG↗

arXiv2026

Finding Stationary Points under Higher-Order Smoothness: Trapping Algorithms and First-Order Lower Bounds

We study the query complexity of finding approximate stationary points under higher-order smoothness using only function values or only gradients. Our results cover different dimensions, oracle models, randomization, and constraints. For constant dimension $d$ and $C^{p,β_p}$ objectives, with $ν:=p+β_p$, we give deterministic function-value gradient trapping algorithms with query complexities $O(\varepsilon^{-(d-1)})$ on $\mathbb R^d$ and $O(\varepsilon^{-(d-1)/ν})$ on $[0,1]^d$. High-order line integration enables gradient-only trapping. For two-dimensional $C^{1,1}$ objectives, randomized integration yields $O(\varepsilon^{-4/3})$ gradient queries on $\mathbb R^2$, breaking the deterministic gradient-only $Θ(\varepsilon^{-2})$ barrier. For general dimension, we introduce a trapping-or-descent framework that lazily reuses boundary models. For $C^{p,β_p}$ objectives with $p\ge2$, it achieves deterministic query complexities $O(d^p+d^{p-1+β_p/(ν-1)} \varepsilon^{-ν/(ν-1)})$ using function values and $O(d^{p-1}+d^{p-2+β_p/(ν-1)} \varepsilon^{-ν/(ν-1)})$ using gradients. These bounds apply to approximate stationary points on $\mathbb R^d$ and approximate KKT points on $[0,1]^d$. Their dimension-independent accuracy exponent matches the classical AR$p$ exponent $(p+1)/p$ when $β_p=1$. For sufficiently high dimension, we prove randomized lower bounds for gradient-only methods. For $C^{1,1}\cap C^{p,1}$ objectives, our lower bound is $Ω(\varepsilon^{-(3/2+1/(2p))})$, matching known upper bounds for $p=2$ and, up to logarithmic factors, for $p=3$. Without the $C^{1,1}$ assumption, we obtain an $Ω(\varepsilon^{-5/3})$ lower bound for every fixed $p\ge3$.

math.OC↗

arXiv2026

Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models

Continual fine-tuning is essential for large language models (LLMs) to dynamically adapt to real-world environments, yet it inevitably suffers from catastrophic forgetting, particularly the performance degradation of previous tasks and LLMs' general-purpose knowledge. Although existing methods, such as orthogonal gradient projection, mitigate the forgetting across various fine-tuning tasks, they fundamentally fail to preserve pre-training LLMs' inherent general-purpose knowledge because the original data and gradients of off-the-shelf pre-training LLMs required by these methods are strictly unknown and highly diverse. To bridge this critical gap, we propose EoupCT, a novel framework designed to Estimate and Orthogonalize Unknown Pre-training gradients for Continual LLM fine-Tuning. Specifically, EoupCT estimates pre-training gradients by dynamically generating pseudo data that is most susceptible to forgetting for new tasks through a learnable soft prompt equipped with Gumbel-Softmax relaxation. Furthermore, we formulate a multi-objective optimization problem and introduce a first-order efficient Pareto optimizer that jointly optimizes LLM parameters and the soft prompt, rigorously enforcing orthogonality between new task updates and the estimated pre-training gradients. Extensive experiments across multiple LLMs demonstrate that EoupCT effectively preserves both task-specific proficiency and inherent general-purpose knowledge, successfully mitigating the catastrophic forgetting.

cs.CL↗

arXiv2026

Bayes-Optimal BER and AUC: Estimation and Evaluation of Estimators

A fundamental quantity in machine learning is the optimal performance achievable by any model on a given task. Estimating this quantity allows us to distinguish the irreducible part of the error from a deficiency of the model, telling us how much room for improvement remains. Recent work has shown that the Bayes error, or equivalently the optimal accuracy, can be estimated from soft labels in binary classification. However, accuracy is often a poor summary of performance in settings with severe class imbalance or noisy annotations, where metrics such as the balanced error rate (BER) and the area under the ROC curve (AUC) are more appropriate. We address this gap with two complementary contributions. (i) Estimation. We propose soft-label-based estimators for the optimal BER and AUC. We first consider the clean setting in which true soft labels and the class prior are known, and then extend the estimators to a more realistic setting in which the class prior is unknown and the observed soft labels are corrupted by an unknown order-preserving transformation, possibly followed by additive noise. In the latter setting, we approximately recover the clean soft labels via isotonic regression with auxiliary hard labels, estimate the class prior with a clipped mean of the hard labels, and derive finite-sample error bounds for the resulting plug-in estimators. (ii) Evaluation. Since the optimum is unobservable on real datasets, evaluating any such estimator is itself nontrivial. We extend the FeeBee framework, originally proposed for evaluating Bayes-error estimators, to the optimal BER and AUC. The resulting procedure provides practical evaluation scores without requiring knowledge of the optimum, and applies to any estimator of the optimal BER or AUC, not only our proposed ones. Experiments on synthetic and real-world datasets validate both the estimators and the evaluation procedure.

cs.LG↗

arXiv2026

Recent advances in weakly supervised learning: New supervision paradigms, assumption relaxations, and practical solutions

Deep learning has achieved great success in recent years thanks to the availability of high-quality, well-annotated training data. However, this requirement is often not met in real-world applications. Weakly supervised learning aims to train an accurate model with incomplete, inexact, or inaccurate supervision. In this chapter, we will discuss recent advances in this field, including new supervision paradigms, relaxed assumptions, and practical solutions. First, we introduce a new weakly supervised binary classification problem called confidence-difference classification and propose consistent approaches to solve it. Next, we investigate complementary-label learning, a weakly supervised multi-class classification problem. Our proposed approaches are based on more relaxed assumptions about the data generation process than existing consistent approaches. Lastly, we present an evaluation framework for partial-label learning, another popular multi-class weakly supervised learning problem, in order to promote fair and realistic evaluation of algorithms in this field.

cs.LG↗

arXiv2026

Proteo-R1: Reasoning Foundation Models for De Novo Protein Design

Deep learning in de novo protein design has achieved atomic-level fidelity. However, existing models remain largely non-deliberative: they directly synthesize molecular geometries without explicitly reasoning about which residues or interactions are functionally essential. As a result, design decisions are entangled with continuous sampling dynamics, limiting interpretability, controllability, and systematic reuse of biochemical knowledge. We introduce Proteo-R1, a reasoning-guided protein design framework that explicitly decouples molecular understanding from geometric generation. Proteo-R1 adopts a dual-expert architecture in which a multimodal large language model (MLLM) serves as an understanding expert, analyzing protein sequences, structures, and textual context to identify key functional residues that govern binding and specificity. These residue-level decisions are then passed as hard constraints to a separate diffusion-based generation expert, which performs conditional co-design while respecting the fixed interaction anchors. This factorization mirrors how human experts approach molecular engineering: first, reasoning about critical interactions, then optimizing geometry subject to those constraints. By operationalizing reasoning as explicit residue-level commitments rather than latent textual guidance, Proteo-R1 achieves stable, interpretable, and modular integration of LLM reasoning with state-of-the-art geometric generative models. Code, data, and demos are available at https://smiles724.github.io/r1/.

cs.LG↗

arXiv2026

When and Why Does Multi-Agent Debate Fail and Does It Really Underperform?

Multi-agent debate (MAD) was proposed as a promising approach for ensembling the wisdom of multiple large language models (LLMs) to improve reasoning and provide effective supervision to superhuman LLMs. However, increasing empirical evidence suggests that MAD may not outperform or even significantly underperform single-agent approaches (SA), raising doubts about the benefits of MAD. In this work, we investigate this issue by analyzing the incentive structures of popular MAD paradigms: (i) competitive MAD (CopMAD) where agents compete by holding opposing positions; (ii) consensus-seeking MAD (CosMAD) where agents are driven to seek consensus. We show that both paradigms suffer from debate hacking: CopMAD reduces to a cheap-talk game, where agents produce misleading messages to win the game, while CosMAD filters out informative disagreements for premature consensus. Consequently, agents in both CopMAD and CosMAD fail to jointly resolve the ambiguity and seek the truth. To this end, we introduce ColMAD, a collaborative protocol that reframes MAD as a non-zero-sum game to encourage agents to provide informative while truthful messages. Through extensive benchmarking on challenging tasks such as error detection, we show that ColMAD significantly outperforms previous MAD protocols up to 10 percentage points. Under the same budgets, ColMAD effectively brings non-trivial improvements over SA methods, implying that the protocol design is critical to realizing the potential of MAD.

cs.LG↗

arXiv2026

Gradient Regularization Mitigates Reward Hacking in Reinforcement Learning from Human Feedback and Verifiable Rewards

Reinforcement Learning from Human Feedback (RLHF) or Verifiable Rewards (RLVR) are two key steps in the post-training of modern Language Models (LMs). A common problem is reward hacking, where the policy may exploit inaccuracies of the reward and learn an unintended behavior. Most previous works address this by limiting the policy update with a Kullback-Leibler (KL) penalty towards a reference model. We propose a different framing: Train the LM in a way that biases policy updates towards regions in which the reward is more accurate. First, we derive a theoretical connection between the accuracy of a reward model and the flatness of an optimum at convergence. Gradient regularization (GR) can then be used to bias training to flatter regions and thereby maintain reward model accuracy. We confirm these results by showing that the gradient norm and reward accuracy are empirically correlated in RLHF. We then empirically show that Reference Resets of the KL penalty find flatter regions with a higher reward accuracy. We further improve on this by proposing to use explicit GR with an efficient finite-difference estimate. Empirically, GR performs better than a KL penalty across a diverse set of RL experiments with LMs. GR achieves a higher GPT-judged win-rate in RLHF, avoids overly focusing on the format in rule-based math rewards, and prevents hacking the judge in LLM-as-a-Judge math tasks.

cs.LG↗