Searcharxiv⌕ Search

arXiv subjects

Se-Young Yun

Publications and source records attributed to Se-Young Yun.

At least 19 recordsLinked to original sources

LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL

Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals. While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static, drawn from fixed corpora. This static approach often results in a critical misalignment between prompt difficulty and policy capability, leaving the judge unable to recover a discriminative reward signal when prompts fail to elicit quality variance among rollouts. To address this misalignment, we introduce LLM-as-a-Tutor, a framework that extends the LLM's role from judge to tutor: a single model serves as an examiner that pairwise-compares policy rollouts to detect non-challenging prompts, and as a generator that appends atomic constraints to them. This append-only design preserves the original prompt while adding constraints to increase difficulty in step with the policy's capability, producing a self-calibrating training signal without external difficulty schedules. On three complex instruction-following benchmarks, our method outperforms both policy-unaware baselines and prior policy-adaptive methods that adapt rubrics or rewrite prompts, on average and on four of five metrics, suggesting prompt adaptation as an underexplored axis of policy-awareness in non-verifiable RL for constraint-heavy instruction following.

cs.AI↗

Variational Monte Carlo for Quantum Excited States via Nested Low-Rank Approximation

The electronic Schrödinger equation encodes a molecule's electronic structure and quantum properties, but solving it scales exponentially with the number of electrons. Solving this equation with neural-network variational Monte Carlo (VMC) has reached chemical accuracy on the ground state of small molecules; extending this success to excited states is constrained by two existing paradigms. Penalty-based methods need $K$ Laplacians per step but require tuning the overlap-penalty weights against the energy gaps they aim to compute. Natural excited states VMC (NES-VMC) is penalty-free, but it combines the $K$ ansatzes into one determinant, whose local energy requires $K^2$ Laplacians per step. We resolve this trade-off via low-rank approximation (LoRA) reparameterization of the rank-$K$ variational objective. The reparameterized objective admits a VMC-compatible stochastic gradient estimator via MCMC sampling from the per-state Born densities, and a sequential nesting scheme recovers the lowest $K$ eigenstates in eigenvalue order. The resulting algorithm, NestedLoRA-VMC, inherits a global optimality guarantee from the LoRA principle. On the first-row atoms Li through Ne at $K=10$, its accuracy is comparable to that of NES-VMC and of a state-of-the-art penalty-based method.

cs.LG↗

Instance-Optimal Estimation with Multiple LLM Judges on a Budget

Evaluating large language models increasingly relies on LLM-as-a-judge protocols, but such evaluations remain costly: different judges have different prices and reliabilities, and the difficulty of each prompt-response pair can vary substantially. This raises a basic allocation question: under a fixed budget, how should one distribute evaluation queries across heterogeneous judges and instances to obtain the most accurate score estimates? We formalize this question as budgeted heteroskedastic multi-judge estimation under independent, unbiased score observations. Given $K$ prompt-response pairs and $J$ judges with known costs but unknown query-judge variances, our goal is to estimate the bounded score vector while minimizing an $\ell_p$-error. Our first contribution is to analyze the inverse-variance weighted estimator (IVWE) and to derive the oracle allocation that minimizes its leading error bound. Since this allocation depends on the unknown variances, we then address the unknown-variance setting by proposing EST-IVWE, an adaptive algorithm that constructs and uses optimistic variance estimates to stabilize the empirical allocation. We prove that EST-IVWE matches the oracle IVWE rate up to lower-order terms in the budget. Our second and central theoretical contribution is a local minimax lower bound matching the leading allocation-dependent rate, establishing asymptotic instance-optimality up to logarithmic factors. An Assouad-type construction based on local, variance-preserving perturbations yields this bound both in expectation and with constant probability. A complementary Fano-type change-of-measure argument captures the logarithmic dependence on the confidence level with a different (suboptimal) allocation geometry. Finally, we numerically validate our approach on synthetic data and resampled HelpSteer2 evaluations, demonstrating improvements over naïve uniform allocation.

cs.LG↗

Bastion: Budget-Aware Speculative Decoding with Tree-structured Block Diffusion Drafting

Block-diffusion drafters have recently emerged as a powerful alternative for speculative decoding by predicting multiple future-token distributions in a single parallel step. However, since these parallel predictions are sampled from position-wise marginals rather than fully conditioned sequences, committing to a single greedy path often fails to capture the target model's preferred trajectory. To address this, we propose BASTION, a budget-aware speculative decoding framework with tree-based diffusion drafting. Unlike existing methods that rely on static tree topologies, BASTION dynamically constructs query-dependent trees by balancing draft quality against hardware constraints. Our framework integrates three synergistic components: (1) an acceptance surrogate that estimates expected accepted length via path confidence, (2) an online latency estimator that calibrates a hardware-aware roofline model, and (3) an adaptive best-first expansion that grows the tree until marginal gains no longer justify incremental verification costs. BASTION is training-free, preserves the target model's distribution, and requires no per-setting tuning. Across diverse benchmarks and GPU architectures, BASTION achieves up to a $6.61\times$ speedup over standard autoregressive decoding, outperforming state-of-the-art block-diffusion baselines by 39%.

cs.LG↗

mSFT: Addressing Dataset Mixtures Overfitting Heterogeneously in Multi-task SFT

Current language model training commonly applies multi-task Supervised Fine-Tuning (SFT) using a homogeneous compute budget across all sub-datasets. This approach is fundamentally sub-optimal: heterogeneous learning dynamics cause faster-learning tasks to overfit early while slower ones remain under-fitted. To address this, we introduce mSFT, an iterative, overfitting-aware search algorithm for multi-task data mixtures. mSFT trains the model on an active mixture, identifies and excludes the earliest overfitting sub-dataset, and reverts to that specific optimal checkpoint before continuing. Extensive evaluations demonstrate that mSFT consistently outperforms 4 baselines across 10 benchmarks and 6 base models. Further analysis confirms mSFT maintains robust gains across diverse dataset sizes, task granularities, and is insensitive to its single new hyperparameter (compute budget). Notably, at low compute budget, mSFT can improve performance while lowering training FLOPs. Ultimately, mSFT establishes a practical overfitting-aware algorithm for multi-task SFT that maximizes the potential of models across diverse data mixtures.

cs.LG↗

A Jointly Efficient and Near-Optimal Algorithm for Heteroskedastic Generalized Linear Bandits with Adversarial Corruptions

We consider heteroskedastic generalized linear bandits (GLBs) with adversarial corruptions and propose \texttt{HCW-GLB-OMD}, a simple algorithm that combines an online mirror descent (OMD)-based estimator with Hessian-based confidence weights. Although both components are well studied, establishing guarantees for the combined algorithm requires a nontrivial analysis. We establish a corruption-robust confidence sequence using technical tools not required in prior analyses, including a weighted likelihood supermartingale and a self-bounding argument. For the regret analysis, we derive a new Hessian potential lemma to handle the interaction between adversarial corruption and adaptive confidence weighting. For self-concordant GLMs, we prove a high-probability regret bound of $\widetilde{\mathcal{O}}\bigl( d \sqrt{\sum_t g(τ_t) \dotμ_{t,\star}} + d^2 g_{\max} κ+ d(1+κ)(C\vee1) \bigr)$, where $\dotμ_{t,\star}$ is the slope of $μ$ at the optimal arm at time $t$, $g(τ_t)$ is the observed dispersion, $g_{\max}=\max_t g(τ_t)$, $κ$ is the worst-case inverse slope of $μ$, and $C$ is the corruption budget. The algorithm attains this bound with one OMD update per round and $\mathcal{O}(d^2)$ storage. We complement this guarantee with matching lower bounds for the leading $\sqrt{T}$ term and, up to $κ$, the $dC$ corruption term.

cs.LG↗

Eliciting Weak-to-Strong Generalization with On-Policy Reverse Distillation

Weak-to-strong generalization asks whether stronger models can learn from weaker supervisors and surpass them. This is particularly important for successive model generations and multi-domain consolidation, where frontier-scale post-training from scratch can be prohibitively expensive. Yet conventional distillation treats the weak teacher as an optimization target, potentially imposing its capacity ceiling on the student. We introduce On-Policy Reverse Distillation (OPRD), which evaluates the teacher's policy shift relative to its reference policy on student rollouts and amplifies the component of the student's verifier-driven policy gradient along that direction. By rescaling only verifier-supported updates, OPRD preserves the stationary points of policy optimization while accelerating learning beyond the teacher. Our evaluations cover mathematical reasoning, logical reasoning, code generation, and visual tool use. In both successive model transfer and multi-teacher distillation, OPRD achieves higher performance with fewer student updates than existing RL and distillation approaches. Response-style analysis shows that OPRD students remain closer to verifier-only RL models than to their weak teachers, suggesting that guidance accelerates rather than redirects student optimization. Strong-to-weak results further show that OPRD combines verifier-driven optimization with teacher guidance regardless of capacity ordering.

cs.LG↗

Language Models Can Control Their Own Attention

Language models spend most of their attention on a small fraction of context, yet they read the entire KV cache to find the few tokens that matter. If the user asks about a previous detail in a 1M-token conversation, global attention layers must scan the full context to generate each token of the reply. A prominent approach pre-selects relevant tokens via lightweight proxy scores, but this extrinsic scoring still incurs O(N) per step. We take an intrinsic approach motivated by the simple question: wouldn't the model already know which parts of the context are relevant? We introduce Declarative Attention (DA), a protocol that elicits the model to declare where it needs to attend within its chain-of-thought, partitioning generation into three modes: (full context), (a specific region), and (recent output only). The inference engine parses these declarations like tool calls and skips most of the KV cache read. Under zero-shot evaluation across 15 long-context tasks, DA on off-the-shelf models (Gemma-4-31B, Qwen-3.6-27B) reduces total attended tokens per response (45.8%, 19.6%), despite extra decoding steps, with modest accuracy drops (1.40pp, 2.10pp) that shrink with model scale. DA unlocks a new axis of sparse attention, with further potential under training-based methods. Code is available at https://github.com/itsnamgyu/declarative-attention.

cs.CL↗

Beyond Second-Order Information: Paid Logistic Regression with Noisy Features

We consider online binary prediction with paid stochastic features, where the learner can spend resources to obtain better observations of the underlying features. This setting is motivated by applications such as medical testing, where additional cost can provide more precise measurements of patient's features for predicting a binary outcome. At each round, a latent feature is drawn i.i.d., the response is generated through a logistic model, and the learner chooses how much to pay before observing a noisy version of the feature. The goal of learner is to jointly control prediction loss and acquisition cost. In contrast to the former result on the online linear regression with paid features, covariance information alone is no longer sufficient: two noise distributions with the same covariance may induce different optimal logistic losses, and lower variance can even lead to higher loss. Under a structural ordering in which larger payments yield better observations by removing independent noise, we obtain $\widetilde O(\sqrt{T})$ regret when the noise structure is known and $\widetilde O(T^{2/3})$ regret when it is unknown, with matching lower bounds up to logarithmic factors.

stat.ML↗

Conversation Is a Two-Body Problem: Dyadic Evaluation of Full-Duplex Dialogue Models

Full-duplex spoken dialogue models listen and speak at the same time, enabling voice agents to have natural, low-latency interactions that turn-based systems cannot offer. However, they are commonly evaluated against single-sided interlocutors: pre-recorded audio that cannot react, or an automated examiner that reacts in real time but only administers a fixed sequence of tests and is never graded. These single-sided frameworks evaluate only half of a two-body problem, where turn-taking, overlap, and interruption are joint products of two coupled speakers. We propose DyaFDB, a framework that evaluates full-duplex models in a dyadic setup: two models converse directly under assigned roles with cooperative or conflicting goals, and both sides are scored offline with an external judge. DyaFDB probes how the two models behave toward each other, such as how they take turns or carry an assigned role under different interests. We instantiate four tasks as 140 scenarios and record 7,560 conversations, covering six self- and cross-play pairings. Throughout the experiments, we observe that how a model behaves continually reshapes its partner. We thus demonstrate that each model must be both the examiner and examinee of the other, and no single fixed interlocutor can play both parts. We will release the scenarios, role prompts, and recording protocols between two full-duplex models, without any pre-recorded audio.

eess.AS↗

Pivot-SD: Efficient Self-Distillation for Masked Diffusion Language Models

Masked diffusion language models (dLMs) offer a promising parallel alternative to autoregressive models for complex reasoning. However, they face a distinct credit-assignment challenge, since a few commitments during denoising sharply reduce the uncertainty over the remaining masked positions and shape much of the response. Most post-training recipes for dLMs do not use this signal to decide which tokens to train on: they typically train on the final text or assign rewards to whole denoising steps, rather than selecting the individual commitments that shape the response. We introduce Pivot-SD, an efficient offline self-distillation framework that supervises only these high-impact commitments (pivots). Pivot-SD selects pivots using an information-gain metric measuring uncertainty reduction over the remaining masked positions. Pivots from successful trajectories are trained with cross-entropy, and pivots from failed trajectories with targeted unlikelihood, leaving the rest of the failed trajectory untouched. Using only 200 questions and four rollouts each, Pivot-SD improves LLaDA-8B-Instruct over full-sequence SFT and budget-matched diffusion RL baselines across math and code benchmarks.

cs.LG↗

Robust Streaming PCA

We consider streaming principal component analysis when the stochastic data generating model is subject to perturbations. While existing models assume a fixed covariance, we adopt a robust perspective where the covariance matrix belongs to a temporal uncertainty set. Under this setting, we provide fundamental limits on convergence of any algorithm recovering principal components. We analyze the convergence of the noisy power method and Oja's algorithm, both studied for the stationary data generating model, and argue that the noisy power method is rate-optimal in our setting. Finally, we demonstrate the validity of our analysis through numerical experiments on synthetic and real-world datasets.

stat.ML↗

PLC-DPO: Posterior Label Correction in Noisy and Ambiguous Preference Optimization

Direct Preference Optimization (DPO) simplifies alignment through pairwise comparisons but assumes all observed preferences are reliable. Real data often violates this assumption, leading to reversed, weak, or ambiguous labels that cause harmful policy updates. To address this, we propose Posterior Label Correction DPO (PLC-DPO) to robustly optimize preferences by routing each pair's training signal as a clean, flip, or tie case. The key idea is to use the calibrated policy-reference margin as online evidence to take appropriate correction actions. This reframes noisy preference learning as actively correcting supervision direction and strength rather than merely filtering suspicious examples. Across 57 dataset-model-benchmark cells, PLC-DPO obtains the best mean win rate against DPO (60.5 vs. 55.5 for the next-best method). Injected-noise and tie stress tests, human disagreement analysis, and self-confirmation diagnostics further show that the routing remains stable and distinguishes flipped from weakly directional pairs.

cs.LG↗

Rethinking Personalized Reward Modeling for LLMs under Preference Heterogeneity via Group-Debiased Federated Learning

Large language models are increasingly aligned to human preferences via reward modeling, but user preference data are sensitive and often cannot be centralized. Federated learning keeps such data local while learning a shared initial reward model, which is later personalized for each client through local fine-tuning. Because users often assign opposite labels to the same pair of responses, existing federated methods address preference heterogeneity by clustering similar clients and training one reward model per group, assuming that each group requires its own initialization. We show that this assumption is unnecessary. Under balanced preference groups, a single FedAvg model, despite starting at nearly random accuracy, surpasses reward models trained separately for each ground-truth group after only a few local optimization steps. We attribute this phenomenon to the flatness of the shared initialization: averaging across all clients learns richer shared representations that distinguish responses while canceling conflicting preference directions, leaving the model near a decision boundary that can be rapidly adapted. Group imbalance breaks this effect as the cancellation becomes asymmetric and leaves minority clients too far from the boundary to recover. Motivated by this observation, we propose FedGD (Federated Learning with Group Debiasing), which discovers latent preference groups during federated training and learns a single reward model using group-debiased client sampling. By counteracting the effect of group imbalance, FedGD learns an initialization that remains highly adaptable, enabling effective personalization without prior knowledge of the underlying groups.

cs.LG↗

Stage-wise Attention-Guided Region Sequencing for Adversarial Attacks on Large Vision-Language Models

Targeted adversarial attacks on Large Vision-Language Models (LVLMs) test whether small image perturbations can steer model responses toward attacker-specified content. Under the standard L-infinity constraint, targeted attacks become a regional perturbation budget allocation problem: attack success depends not only on the perturbation objective, but also on which regions receive updates and in what order. Existing localized attacks improve over global perturbations but rely on stochastic spatial sampling, often updating weakly influential regions. We address this limitation through an attention-based analysis showing that cross-modal attention identifies adversarially sensitive regions and that perturbing high-attention hotspots induces predictable redistribution toward subsequent salient regions. These findings motivate attention-guided region sequencing, which begins from dominant hotspots and progressively moves the update support toward next-salient regions. Based on these principles, we propose Stage-wise Attention-Guided Attack (SAGA), a black-box region-sequencing framework that uses a fixed attention map from an open-source LVLM to guide perturbation updates without accessing target-model parameters, gradients, or attention maps. Across ten closed-source and open-source LVLMs, SAGA achieves state-of-the-art attack success rates and the best overall imperceptibility. The source code is available at https://github.com/jaehyun-kwak/SAGA.

cs.CV↗

Predicting LLM Reasoning Performance with Small Proxy Model

Given the prohibitive cost of pre-training large language models, it is essential to leverage smaller proxy models to optimize datasets before scaling up. However, this approach becomes challenging for reasoning capabilities, which exhibit emergent behavior that only appear reliably at larger model sizes, often exceeding 7B parameters. To address this, we introduce rBridge, showing that small proxies ($\leq$1B) can effectively predict large-model reasoning by aligning more closely with (1) the pre-training objective and (2) the target task. rBridge achieves this by weighting negative log-likelihood with task alignment, using reasoning traces from frontier models as gold labels. In our experiments, rBridge (i) reduces dataset ranking costs by over 100x relative to the best baseline, (ii) achieves the strongest correlation across six reasoning benchmarks at 1B to 32B scale, and (iii) zero-shot transfers predictive relationships across pre-training datasets at 1B to 7B scale. These findings indicate that rBridge offers a practical path for exploring reasoning-oriented pre-training at lower cost.

cs.LG↗

CascadeFormer: Depth-Tapered Transformers Motivated by Gradient Fan-in Asymmetry

Deep Transformers are composed of uniformly stacked residual blocks, yet their deepest layers often add little value. We present two efficiency methods that exploit this asymmetry. CascadeFormer tapers width with depth to match the uneven information flow across layers, achieving comparable perplexity to a uniform baseline at the same training budget while reducing latency by 8.6% and increasing throughput by 9.4%. CascadeFlow Pruning removes layers using accumulated training gradients, with no post hoc analysis. It outperforms standard heuristics on perplexity and rank-stability and stays competitive on downstream accuracy. To motivate these methods, we propose Gradient Fan-in Asymmetry (GFA) as a structural account of why deeper layers contribute less. In Pre-LayerNorm residual stacks, the gradient at a layer is the sum of an identity path and all downstream functional paths, producing a gradient fan-in that decays linearly with depth (and quadratically under deep supervision), yielding richer gradients for early layers and sparser ones for later layers. We provide correlational and interventional evidence for GFA on models trained from scratch up to 1.2B parameters. Across Transformers and ResNets, accumulated training gradients follow the theoretical fan-in and are associated with post hoc layer importance. Two interventions point to structure rather than magnitude as the bottleneck: equalizing per-layer gradient norms does not restore late-layer value, while increasing downstream path counts via parameter-shared repetition restores and elevates it. Whether gradient magnitude proxies fan-in beyond high-rank regimes, and how these dynamics behave at the 100B+ scale, remain open questions.

cs.LG↗

SpotAttention: Plug-In Block-Sparse Routing for Pretrained Long-Context Transformers

Long contexts have become standard in pretrained LLMs, yet they remain expensive to run: prefill compute grows quadratically with sequence length, and every decode step re-reads a key-value cache that grows linearly with it. Sparse attention cuts these costs by attending only to a relevant subset of past tokens, but selecting that subset is itself expensive. We present SpotAttention, a lightweight selector that attaches to a frozen pretrained transformer and learns by KL distillation to estimate its attention distribution. The selector picks the top-K keys each query attends to, and because its estimate is a calibrated distribution, a dual top-p rule reads the per-query, per-layer budget directly from it. Across Qwen3 (dense, 4B-32B) and Qwen3.5 (hybrid linear/full attention, 4B-9B), SpotAttention matches dense accuracy at contexts up to 128K tokens, eight times the training length. Decode at L=128K runs 3.9x faster than FlashAttention and 1.8x faster than Twilight, the strongest training-free baseline. Quantizing the selector's K-cache to INT4 or FP4 microscale shrinks it 3.5x at no accuracy cost.

cs.LG↗