arXiv subjects

Hongyi Wen

Publications and source records attributed to Hongyi Wen.

At least 19 recordsLinked to original sources

arXiv2026

H$^2$R: Hierarchical Hint-Based Reward for Mathematical Reasoning via Reinforcement Learning

Reinforcement Learning with Verifiable Rewards has shown great promise in enhancing mathematical reasoning in large language models. However, existing approaches provide limited intermediate supervision for complex, multi-step reasoning tasks. To bridge this gap, we propose \textbf{Hierarchical Hint-based Reward}, a framework that leverages the intrinsic sub-question structure of mathematical problems as structured intermediate conditioning signals during RLVR training. Unlike chain-of-thought prompting, these hierarchical hints serve as logically grounded intermediate milestones that effectively reduce the reasoning horizon toward the final answer. To validate this framework, we introduce \textbf{HiMath}, a benchmark of hierarchically structured mathematical problems. Across diverse LLMs, hierarchical hints consistently improve reasoning performance even without task-specific training. When integrated into RL training via H$^2$R, they boost Qwen2.5-Math-7B accuracy on HiMath from 30.4\% to 39.1\% using hierarchical hints, and up to 45.6\% with Oracle Hierarchical Hints. This pronounced improvement demonstrates that incorporating reliable intermediate checkpoints into reward design is essential for sustaining complex mathematical reasoning.

cs.LG↗

arXiv2026

Learning to Be Guided: Shaping Latent Geometry for Classifier-Free Guidance

Classifier-free guidance (CFG) is a standard component of conditional diffusion models, and guided generation performance has become the \textit{de facto} measure of state-of-the-art models. Recent work on latent diffusion models seeks to learn latent spaces that are easier to model with diffusion. However, under CFG, generation additionally depends on how conditions are arranged in the latent space. We study this arrangement under discrete class conditions, termed \emph{class geometry}, and investigate what class geometry allows a latent space to work well with CFG. Under a Gaussian class model, we show that shifting variance from intra- to inter-class at fixed total variance makes classes more distinguishable under noise, and derive a noisy posterior objective whose per-sample descent direction is proportional to the oracle CFG correction. Based on the analysis, we propose Prototype Regularization (PR), a lightweight surrogate for the intractable noisy posterior objective, which applies a cosine prototype classifier to spatially pooled latents. With a controlled training budget (20 epochs for VAE, 80 epochs for DiT), adding PR to an LDM baseline reduces guided gFID on ImageNet from 3.31 to 2.28, surpassing a DINOv3-aligned VA-VAE (2.44) with negligible overhead. We further demonstrate that PR generalizes to pseudo-labels derived from clustering, continuous text-embedding conditions, and video latents. Together, our findings highlight class geometry as a critical factor in guided generation and take a first step toward CFG-compatible representation learning. Our project page is available at: https://maps-research.github.io/cfg-geometry/.

cs.LG↗

arXiv2026

Distributionally Robust Mixture-of-Experts Training

Mixture-of-Experts (MoE) transformers scale capacity by activating only a few experts per token, but this sparsity creates a hidden reliability problem: when routing is imperfect, load-balanced models may send tokens to experts that are insufficiently trained for the assigned inputs. We propose Distributionally Robust MoE Training (DRMoET), a drop-in objective that treats layer-wise experts as endogenous robustness groups and optimizes high-loss routing outcomes rather than merely equalizing traffic. DRMoET updates a per-layer expert distribution by an entropy-regularized softmax rule on EMA-smoothed, activation-weighted expert losses, strengthening plausible non-top routing paths while preserving standard MoE computation. Under the FLAME-MoE recipe at 746M-total and 10.3B-total scales, DRMoET improves downstream averages over both standard FLAME-MoE and auxiliary-loss-free balancing. At 10.3B total parameters and 67B training tokens, DRMoET improves the seven-task average from 0.6625 to 0.6767, while the auxiliary-loss-free baseline achieves 0.6431. Mechanistic analyses show lower expert-loss variance with nearly unchanged mean loss, 4.3% lower excess loss under forced mid-$k$ misrouting, and improved domain-expert specialization. These results position routing robustness-not only utilization balance-as a practical objective for reliable sparse MoE scaling. Project page and code are available at: https://drmoet.github.io/.

cs.LG↗

arXiv2026

Coherence-Aware Distributional Evaluation of Open-Ended Text Generation

Existing open-ended generation metrics measure likelihood, lexical diversity, or distributional similarity in generic representation space, yet can miss fundamental dimensions of quality. A prominent blind spot is global coherence: a generated passage may be locally fluent while remaining globally contradictory, causally inconsistent, or topically disconnected. We identify representation as a central bottleneck in detecting these failures and introduce CHORD (Coherence-aware Hidden-state Open-generation Reference Distance), a coherence-sensitive distributional metric. CHORD encodes generated and human-written corpora in the hidden-state space of a frozen LLM using a coherence-eliciting prompt, and compares the resulting distributions using RBF-MMD. To test coherence sensitivity and selectivity, we construct a counterfactual evaluation suite pairing graded coherence-degrading perturbations with meaning-preserving controls. CHORD selectively detects relation, discourse, structural, and mixture failures that perplexity, entropy, MAUVE, FBD, and MMD-based baselines either miss or cannot separate from benign rewriting. Factorial ablations show that representation is the primary source of coherence sensitivity, while RBF-MMD improves sample efficiency. Larger backbones capture finer-grained distinctions, but coherence prompting improves selectivity only when the backbone can follow the prompt. On unconditional generation and prefix continuation, CHORD yields model rankings that strongly align with human judgments of whether outputs make sense and appear human-written. Together, these results establish representation design as central to reliable distributional evaluation. Code: https://github.com/MAPS-research/CHORD. Experiments: https://github.com/MAPS-research/CHORD-Experiment.

cs.CL↗

arXiv2026

NovelAPIBench: Diagnosing How A Code LLM Learns to Use Novel APIs

Rapidly evolving software libraries require coding language models to use unfamiliar APIs. Yet failed solutions alone tell us little about what API knowledge models need or how to provide it. We present NovelAPIBench, an automated diagnostic framework that builds benchmarks around the API knowledge gaps of a target model. It identifies novel APIs, generates executable coding tasks with separately controllable knowledge components, and classifies failures into six categories. The framework can be rerun as models and libraries evolve. We study 21 libraries across five domains and six backbone models, with our primary benchmark comprising 1,670 tasks covering 856 APIs. Our experiments show that when supplied directly, usage examples provide the strongest standalone guidance for novel API acquisition, and combining them with signature descriptions approaches full-knowledge performance. Retrieving this knowledge from a pool of APIs yields lower accuracy than supplying only the target API's knowledge, even when retrieval finds the target. Fine-tuning on other APIs improves performance on unseen APIs mainly when external knowledge is supplied, helping models both select the target API and integrate it into the surrounding code. These findings highlight complementary roles for external API knowledge and the ability to apply it. By separating knowledge content, delivery, and use, NovelAPIBench helps diagnose what limits novel API use and guides adaptation to evolving libraries. Code and data are available at https://github.com/MAPS-research/NovelAPIBench.

cs.AI↗

arXiv2026

EvoGenUI-Bench: Evaluating LLMs as Multi-Turn Generative UI Assistants

Large language models can generate interactive web interfaces, but reliable generative UI requires maintaining an executable artifact as user requests evolve. We introduce EvoGenUI-Bench, a benchmark for multi-turn interface maintenance comprising 150 five-turn tasks and 750 turns across three scenarios: information presentation, executable interaction, and tool-grounded external state. We execute generated artifacts in a browser and evaluate them using screenshots, source and DOM evidence, actor traces, and runtime logs. Beyond turn-level and episode-level success, we measure cross-turn retention with Adjacent Pass Retention. Across eight models, even the strongest achieves 74.9% Turn Pass while completing only 37.3% of five-turn episodes; APR further falls to 52.4% on tool-grounded tasks. Diagnostic analysis shows that presentation failures center on information architecture, interaction failures on derived-state propagation and affordance binding, and tool-grounded failures additionally involve external-state grounding and requirement decomposition. These results reframe generative UI evaluation from judging isolated outputs to testing whether interface behavior, derived state, external state, and assistant claims remain synchronized as the artifact evolves.

cs.AI↗

arXiv2026

Toward Latent Language Model Skills Steering and Optimization: An Empirical Study

Skills, as a useful abstraction for the procedural capabilities of large language models (LLMs), capture how models perform structured, multi-step reasoning and program execution. Existing approaches typically treat skills as explicit, surface-level constructs specified through prompts or programs, leaving open the question of how such procedural capabilities are represented inside the model and whether they can be manipulated as structured objects in latent space. In this empirical study, we investigate whether procedural LLM skills can be represented as directions in activation space and whether vector-space operations over these directions can express skill-level behaviors. We find that procedural skills admit a vector-space representation: individual skill directions can be activated to shift model behavior; independently extracted directions can compose to form higher-level skills. Contrastive directions yield context-conditioned algorithmic personalization and optimization trajectories over skill directions evolve non-monotonically, with intermediate states often surpassing fully optimized solutions. These results support a representation-level view of procedural LLM skills: they admit a latent vector-space organization that allows direct manipulation through internal interventions.

cs.AI↗

arXiv2026

Histogram-constrained Image Generation

Diffusion models have emerged as a dominant paradigm in generative modeling, enabling high-fidelity sampling from complex data distributions. Despite impressive capabilities, controlling diffusion models to produce outputs aligned with user intent remains an open challenge, especially when balancing global coherence with local precision. Existing control mechanisms vary in the granularity of their conditioning signals. For example, textual prompts guide generation globally through high-level semantics, while ControlNet-like approaches secure precise local structure via dense conditions. In this work, we introduce Histogram-constrained Image Generation (HIG), a novel control mechanism that falls into the middle ground of control granularity. Our framework enforces user-specified distributional constraints (e.g., color histograms or latent token distributions) during the generation process with exact precision. We model such control as an optimal transport (OT) problem and apply explicit guidance transformations during sampling, thereby driving the diffusion trajectory to align with the desired histogram. We demonstrate the versatility of HIG across diverse applications, including constrained generation via color/latent histograms and high-capacity information embedding through histogram-level encoding. Our findings underscore the promise of distributional control, a flexible and interpretable control scheme that is fully compatible with existing control mechanisms, diversifying the hybrid strategies for controllable image generation. Our project page is available at: https://maps-research.github.io/hig/.

cs.CV↗

arXiv2026

Evaluating Interactivity: Toward Automated Assessment of AI-Generated Explorable Explanations

While large language models now enable rapid generation of interactive learning materials, evaluating the interaction quality of these explorable explanations remains an open challenge. Existing benchmarks largely focus on code executability or visual fidelity, providing limited insight into dynamic interaction behaviors such as learner-controlled state transitions and context-sensitive system responses, which are factors that critically shape learners' conceptual understanding. We present EE-Eval, an automated evaluation framework that formalizes interactivity as a finite space of learner-controllable states and transitions, represented as a Finite State Machine (FSM). By extracting FSMs from AI-generated explorable explanations, EE-Eval externalizes implicit interaction logic into an explicit, machine-interpretable graph. Evaluation is performed by comparing each generated FSM to an ideal FSM that encodes pedagogical intent, using a combination of graph-based metrics and embedding-based comparison of states, actions, and feedback to measure their structural and semantic similarity. Across thousands of generated explorable explanations spanning 127 concepts and produced by 6 AI models, EE-Eval consistently differentiates interaction quality beyond surface-level criteria such as functional correctness or visual quality, and exhibits substantially stronger alignment with human judgments of interactivity and pedagogical effectiveness than existing baselines. By framing interactivity as testable behavioral models rather than an emergent byproduct of LLM generation, EE-Eval transforms evaluation into a reflective diagnostic tool, enabling pedagogically grounded and actionable human-AI collaboration in creating interactive educational content.

cs.HC↗

arXiv2026

SHAPE: Unifying Safety, Helpfulness and Pedagogy for Educational LLMs

Large Language Models (LLMs) have been widely explored in educational scenarios. We identify a critical vulnerability in current educational LLMs, pedagogical jailbreaks, where students use answer-inducing prompts to elicit solutions rather than scaffolded instructions. To enable systematic study, we unify and formalize safe, helpful, and pedagogical behaviors with a knowledge-mastery graph and introduce SHAPE, a benchmark of 9,087 student-question pairs for evaluating tutoring behavior under adversarial pressure. We propose a graph-augmented tutoring pipeline that infers prerequisite concepts from queries, identifies mastery gaps, and routes generation between instructing and problem-solving via explicit gating. Experiments across multiple LLMs show that our method yields significantly improved safety under two pedagogical jailbreak settings, while maintaining near-ceiling helpfulness under the same evaluation protocol. Our code and data are available at https://github.com/MAPS-research/SHaPE

cs.CL↗

arXiv2026

From Navigation to Refinement: Revealing the Two-Stage Nature of Flow-based Diffusion Models through Oracle Velocity

Flow-based diffusion models have emerged as a leading paradigm for training generative models across images and videos. However, their memorization-generalization behavior remains poorly understood. In this work, we revisit the flow matching (FM) objective and study its marginal velocity field, which admits a closed-form expression, allowing exact computation of the oracle FM target. Analyzing this oracle velocity field reveals that flow-based diffusion models inherently formulate a two-stage training target: an early stage guided by a mixture of data modes, and a later stage dominated by the nearest data sample. The two-stage objective leads to distinct learning behaviors: the early navigation stage generalizes across data modes to form global layouts, whereas the later refinement stage increasingly memorizes fine-grained details. Leveraging these insights, we explain the effectiveness of practical techniques such as timestep-shifted schedules, classifier-free guidance intervals, and latent space design choices. Our study deepens the understanding of diffusion model training dynamics and offers principles for guiding future architectural and algorithmic improvements. Our project page is available at: https://maps-research.github.io/from-navigation-to-refinement/.

cs.LG↗

arXiv2025

An Analysis of Large Language Models for Simulating User Responses in Surveys

Using Large Language Models (LLMs) to simulate user opinions has received growing attention. Yet LLMs, especially trained with reinforcement learning from human feedback (RLHF), are known to exhibit biases toward dominant viewpoints, raising concerns about their ability to represent users from diverse demographic and cultural backgrounds. In this work, we examine the extent to which LLMs can simulate human responses to cross-domain survey questions through direct prompting and chain-of-thought prompting. We further propose a claim diversification method CLAIMSIM, which elicits viewpoints from LLM parametric knowledge as contextual input. Experiments on the survey question answering task indicate that, while CLAIMSIM produces more diverse responses, both approaches struggle to accurately simulate users. Further analysis reveals two key limitations: (1) LLMs tend to maintain fixed viewpoints across varying demographic features, and generate single-perspective claims; and (2) when presented with conflicting claims, LLMs struggle to reason over nuanced differences among demographic features, limiting their ability to adapt responses to specific user profiles.

cs.CL↗

arXiv2025

ImageGem: In-the-wild Generative Image Interaction Dataset for Generative Model Personalization

We introduce ImageGem, a dataset for studying generative models that understand fine-grained individual preferences. We posit that a key challenge hindering the development of such a generative model is the lack of in-the-wild and fine-grained user preference annotations. Our dataset features real-world interaction data from 57K users, who collectively have built 242K customized LoRAs, written 3M text prompts, and created 5M generated images. With user preference annotations from our dataset, we were able to train better preference alignment models. In addition, leveraging individual user preference, we investigated the performance of retrieval models and a vision-language model on personalized image retrieval and generative model recommendation. Finally, we propose an end-to-end framework for editing customized diffusion models in a latent weight space to align with individual user preferences. Our results demonstrate that the ImageGem dataset enables, for the first time, a new paradigm for generative model personalization.

cs.CV↗

arXiv2025

Reveal and Release: Iterative LLM Unlearning with Self-generated Data

Large language model (LLM) unlearning has demonstrated effectiveness in removing the influence of undesirable data (also known as forget data). Existing approaches typically assume full access to the forget dataset, overlooking two key challenges: (1) Forget data is often privacy-sensitive, rare, or legally regulated, making it expensive or impractical to obtain (2) The distribution of available forget data may not align with how that information is represented within the model. To address these limitations, we propose a ``Reveal-and-Release'' method to unlearn with self-generated data, where we prompt the model to reveal what it knows using optimized instructions. To fully utilize the self-generated forget data, we propose an iterative unlearning framework, where we make incremental adjustments to the model's weight space with parameter-efficient modules trained on the forget data. Experimental results demonstrate that our method balances the tradeoff between forget quality and utility preservation.

cs.CL↗

arXiv2024

Diffusion Cocktail: Mixing Domain-Specific Diffusion Models for Diversified Image Generations

Diffusion models, capable of high-quality image generation, receive unparalleled popularity for their ease of extension. Active users have created a massive collection of domain-specific diffusion models by fine-tuning base models on self-collected datasets. Recent work has focused on improving a single diffusion model by uncovering semantic and visual information encoded in various architecture components. However, those methods overlook the vastly available set of fine-tuned diffusion models and, therefore, miss the opportunity to utilize their combined capacity for novel generation. In this work, we propose Diffusion Cocktail (Ditail), a training-free method that transfers style and content information between multiple diffusion models. This allows us to perform diversified generations using a set of diffusion models, resulting in novel images unobtainable by a single model. Ditail also offers fine-grained control of the generation process, which enables flexible manipulations of styles and contents. With these properties, Ditail excels in numerous applications, including style transfer guided by diffusion models, novel-style image generation, and image manipulation via prompts or collage inputs.

cs.CV↗

arXiv2023

GEMRec: Towards Generative Model Recommendation

Recommender Systems are built to retrieve relevant items to satisfy users' information needs. The candidate corpus usually consists of a finite set of items that are ready to be served, such as videos, products, or articles. With recent advances in Generative AI such as GPT and Diffusion models, a new form of recommendation task is yet to be explored where items are to be created by generative models with personalized prompts. Taking image generation as an example, with a single prompt from the user and access to a generative model, it is possible to generate hundreds of new images in a few minutes. How shall we attain personalization in the presence of "infinite" items? In this preliminary study, we propose a two-stage framework, namely Prompt-Model Retrieval and Generated Item Ranking, to approach this new task formulation. We release GEMRec-18K, a prompt-model interaction dataset with 18K images generated by 200 publicly-available generative models paired with a diverse set of 90 textual prompts. Our findings demonstrate the promise of generative model recommendation as a novel personalization problem and the limitations of existing evaluation metrics. We highlight future directions for the RecSys community to advance towards generative recommender systems. Our code and dataset are available at https://github.com/MAPS-research/GEMRec.

cs.IR↗

arXiv2021

Evaluating Music Recommendations with Binary Feedback for Multiple Stakeholders

High quality user feedback data is essential to training and evaluating a successful music recommendation system, particularly one that has to balance the needs of multiple stakeholders. Most existing music datasets suffer from noisy feedback and self-selection biases inherent in the data collected by music platforms. Using the Piki Music dataset of 500k ratings collected over a two-year time period, we evaluate the performance of classic recommendation algorithms on three important stakeholders: consumers, well-known artists and lesser-known artists. We show that a matrix factorization algorithm trained on both likes and dislikes performs significantly better compared to one trained only on likes for all three stakeholders.

cs.IR↗

arXiv2021

Correcting the User Feedback-Loop Bias for Recommendation Systems

Selection bias is prevalent in the data for training and evaluating recommendation systems with explicit feedback. For example, users tend to rate items they like. However, when rating an item concerning a specific user, most of the recommendation algorithms tend to rely too much on his/her rating (feedback) history. This introduces implicit bias on the recommendation system, which is referred to as user feedback-loop bias in this paper. We propose a systematic and dynamic way to correct such bias and to obtain more diverse and objective recommendations by utilizing temporal rating information. Specifically, our method includes a deep-learning component to learn each user's dynamic rating history embedding for the estimation of the probability distribution of the items that the user rates sequentially. These estimated dynamic exposure probabilities are then used as propensity scores to train an inverse-propensity-scoring (IPS) rating predictor. We empirically validated the existence of such user feedback-loop bias in real world recommendation systems and compared the performance of our method with the baseline models that are either without de-biasing or with propensity scores estimated by other methods. The results show the superiority of our approach.

cs.IR↗