arXiv subjects

Siyuan Song

Publications and source records attributed to Siyuan Song.

At least 19 recordsLinked to original sources

arXiv2026

Reasoning or Memorization: Can LLMs Understand and Generate Chinese Xiehouyu Riddles?

In this paper, we push the boundary of LLM reasoning by testing them in a Chinese language game, xiehouyu, with novel xiehouyu created by linguists that had not existed before to avoid data contamination. We use multiple-choice questions (MCQ), free-form explanation generation, and new xiehouyu creation to evaluate LLMs' ability to understand and create xiehouyu. In MCQ, we use the delta of accuracy ($\Delta_{acc}$) between existing but low-frequency xiehouyu and novel ones as an index for memorization. $\Delta_{acc}$ for native speakers is very low, suggesting similar processing mechanisms. However, we found that frontier Chinese models have on average a $\Delta_{acc}$ of 23.6\%, while English-centric models tested have a mean $\Delta_{acc}$ of 5.1\%, suggesting that frontier Chinese models are likely trained with much larger Chinese data, thus memorizing more low-frequency xiehouyu. For novel xiehouyu, Gemini 3.1 Pro demonstrated remarkable ability with acc 92.6, which is 24\% higher than human accuracy. In xiehouyu creation, those created by LLMs receive much worse ratings than those by humans. These results suggest that claims about the reasoning abilities of LLMs may need careful re-examination considering the data contamination issue, and that LLMs' creativity in language-related tasks may still be behind human experts, at least in Chinese xiehouyu.

cs.CL↗

arXiv2026

The First ChineseBabyLM Challenge: training data-efficient and cognitively plausible language models for Chinese

This paper presents the first ChineseBabyLM Challenge, organized as part of NLPCC 2026. The challenge asked participants to train language models from scratch using no more than 102M Chinese words. The models were evaluated on three tracks: natural language understanding, cognitive alignment, and Hanzi knowledge. There were no restrictions on tokenizers, model architectures, or the number of training epochs. Eighteen teams submitted 28 distinct models, generating 74 result files. The overall-winning team used a DeBERTa-v2 architecture and introduced an auxiliary pinyin-prediction objective during pretraining. Several submissions also explored curriculum-learning strategies and architectural innovations. Overall, the challenge provides a benchmark for advancing data-efficient and cognitively plausible approaches to Chinese language modeling.

cs.CL↗

arXiv2026

GR2 Technical Report

Industrial recommendation systems serve billions of users through a multi-stage funnel -- retrieval, early-stage ranking, and re-ranking -- where the final re-ranking step disproportionately shapes user engagement and downstream performance, particularly for carousel and grid display formats. Despite growing enthusiasm for Large Language Models (LLMs) in recommendation, three gaps hinder industrial adoption: (1) most efforts target retrieval and ranking, leaving re-ranking -- the stage closest to the final user experience -- largely underexplored; (2) LLMs are typically deployed zero-shot or via supervised fine-tuning, underutilizing the reasoning capabilities unlocked by reinforcement learning (RL) on verifiable rewards; (3) deployed catalogs index billions of items with non-semantic identifiers that lie outside any base-LLM vocabulary. We present GR2 (Generative Reasoning Re-Ranker), an end-to-end framework that combines (i) mid-training on semantic IDs produced by a tokenizer with >=99% uniqueness, (ii) reasoning-trace distilled from a stronger teacher via targeted prompting and rejection sampling, and (iii) RL with verifiable rewards purpose-built for re-ranking. To make GR2 resource-viable, we further (iv) introduce a context compressor that amortizes training cost, On-Policy Distillation (OPD) as a scalable alternative to SFT -- which we find collapses at industrial scale -- and reasoning distillation for low-latency serving. GR2 delivers +18.7% R@1, +7.1% R@3, and +9.6% N@3 over legacy baselines on industrial-scale traffic. We further find that reward design is critical in re-ranking: LLMs often hack rewards by preserving the incoming order or exploiting position bias, motivating conditional verifiable rewards as essential industrial components.

cs.IR↗

arXiv2026

RoamFlow: Reinforcement-Aligned One-Step Action MeanFlow Policy for Image-Goal Navigation

Image-goal navigation is a key challenge in embodied robotics, where an agent must reach a target specified solely by a goal image. While existing reinforcement learning approaches map perceptual observations directly to actions, they struggle to model long-horizon dependencies, often leading to suboptimal trajectories. To address this limitation, we propose RoamFlow, a generative navigation framework that leverages MeanFlow to predict the average velocity field for trajectory synthesis, enabling efficient few-step generation and reducing inference latency. We further adopt a two-stage training strategy that combines expert imitation for stable initialization with reinforcement learning for task-specific policy refinement. Extensive experiments in both Habitat simulation and real-world robotic platforms demonstrate that RoamFlow achieves efficient inference while maintaining strong navigation performance under real-time constraints.

cs.RO↗

arXiv2026

DeTrack: A Benchmark and Altitude-Aware Dual World Model for Drone-embodied Tracking

Aerial object tracking has broad applications in public safety, emergency rescue, wildlife monitoring, and related fields. However, existing aerial tracking benchmarks are mainly based on passive 2D video sequences captured from fixed camera locations or predefined flight paths, where drones are treated as passive cameras rather than embodied agents that actively perceive, interact, and control their motion in dynamic 3D scenes. In this paper, we define a new drone-embodied tracking task, termed DeTrack, which requires a drone to track a target in interactive 3D environments using online egocentric observations and active flight control in a closed loop. We build a large-scale benchmark containing 11,368 target trajectories across diverse scenes, rendering conditions, semantic regions, and moving distractors, together with evaluation metrics for target visibility, tracking accuracy, and trajectory success. We further propose AaDWorlds, an altitude-aware dual world model framework for drone-embodied tracking. AaDWorlds consists of an altitude-aware perception module and dual world models that imagine future states under both high- and low-altitude regimes. By combining pseudo altitude-aware observations and imagined future states, AaDWorlds alleviates the intrinsic altitude-mediated contradiction between target visibility and flight safety. Experiments on the DeTrack benchmark demonstrate that AaDWorlds improves closed-loop tracking performance across all evaluation metrics.

cs.CV↗

arXiv2026

Structure and Progress Aware Diffusion for Medical Image Segmentation

Medical image segmentation is crucial for computer-aided diagnosis, which necessitates understanding both coarse morphological and semantic structures, as well as carving fine boundaries. The morphological and semantic structures in medical images are beneficial and stable clues for target understanding. While the fine boundaries of medical targets (like tumors and lesions) are usually ambiguous and noisy since lesion overlap, annotation uncertainty, and so on, making it not reliable to serve as early supervision. However, existing methods simultaneously learn coarse structures and fine boundaries throughout the training process. In this paper, we propose a structure and progress-aware diffusion (SPAD) for medical image segmentation, which consists of a semantic-concentrated diffusion (ScD) and a boundary-centralized diffusion (BcD) modulated by a progress-aware scheduler (PaS). Specifically, the semantic-concentrated diffusion introduces anchor-preserved target perturbation, which perturbs pixels within a medical target but preserves unaltered areas as semantic anchors, encouraging the model to infer noisy target areas from the surrounding semantic context. The boundary-centralized diffusion introduces progress-aware boundary noise, which blurs unreliable and ambiguous boundaries, thus compelling the model to focus on coarse but stable anatomical morphology and global semantics. Furthermore, the progress-aware scheduler gradually modulates noise intensity of the ScD and BcD forming a coarse-to-fine diffusion paradigm, which encourage focusing on coarse morphological and semantic structures during early target understanding stages and gradually shifting to fine target boundaries during later contour adjusting stages.

cs.CV↗

arXiv2026

A Very Big Video Reasoning Suite

Rapid progress in video models has largely focused on visual quality, leaving their reasoning capabilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale training data. To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks following a principled taxonomy and over one million video clips, approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark toolkit, and models are publicly available at https://video-reason.com/?v=vbvr .

cs.CV↗

arXiv2026

Bears, all bears, and some bears. Language Constraints on Language Models' Inductive Inferences

Language places subtle constraints on how we make inductive inferences. Developmental evidence by Gelman et al. (2002) has shown children (4 years and older) to differentiate among generic statements ("Bears are daxable"), universally quantified NPs ("all bears are daxable") and indefinite plural NPs ("some bears are daxable") in extending novel properties to a specific member (all > generics > some), suggesting that they represent these types of propositions differently. We test if these subtle differences arise in general purpose statistical learners like Vision Language Models, by replicating the original experiment. On tasking them through a series of precondition tests (robust identification of categories in images and sensitivities to all and some), followed by the original experiment, we find behavioral alignment between models and humans. Post-hoc analyses on their representations revealed that these differences are organized based on inductive constraints and not surface-form differences.

cs.CL↗

arXiv2026

All-optical photoacoustic tomography via beam deflection

Photoacoustic imaging (PAI) uniquely combines the advantages of optical contrast with deep tissue penetration capability of acoustic waves, enabling imaging at depths of several centimeters. Conventional photoacoustic imaging methods have relied on pulsed lasers to induce the photoacoustic effect, coupled with arrays of pressure transducers to detect the resulting ultrasound signals. In this work, we propose an alternative all-optical approach that leverages optical deflection to record photoacoustic waves by an array of detection beams. The measured signal is shown to be the Radon transform of the pressure gradients. An optimization-based inversion procedure is used to reconstruct the initial time pressure gradient field. Subsequently, a Galerkin method is used to reconstruct the pressure field from the pressure gradient field. The new modality offers the potential for enhanced sensitivity and reduced signal distortion, advancing the capabilities of photoacoustic imaging beyond traditional transducer-based systems.

physics.app-ph↗

arXiv2025

Development of the CEPC analog hadron calorimeter prototype

The Circular Electron Positron Collider (CEPC) is a next-generation electron$-$positron collider proposed for the precise measurement of the properties of the Higgs boson. To emphasize boson separation and jet reconstruction, the baseline design of the CEPC detector was guided by the particle flow algorithm (PFA) concept. As one of the calorimeter options, the analogue hadron calorimeter (AHCAL) was proposed. The CEPC AHCAL comprises a 40-layer sandwich structure using steel plates as absorbers and scintillator tiles coupled with silicon photomultipliers (SiPM) as sensitive units. To validate the feasibility of the AHCAL option, a series of studies were conducted to develop a prototype. This AHCAL prototype underwent an electronic test and a cosmic ray test to assess its performance and ensure it was ready for three beam tests performed in 2022 and 2023. The test beam data is currently under analysis, and the results are expected to deepen our understanding of hadron showers, validate the concept of Particle Flow Algorithm (PFA), and ultimately refine the design of the CEPC detector.

physics.ins-det↗

arXiv2025

What Can String Probability Tell Us About Grammaticality?

What have language models (LMs) learned about grammar? This question remains hotly debated, with major ramifications for linguistic theory. However, since probability and grammaticality are distinct notions in linguistics, it is not obvious what string probabilities can reveal about an LM's underlying grammatical knowledge. We present a theoretical analysis of the relationship between grammar, meaning, and string probability, based on simple assumptions about the generative process of corpus data. Our framework makes three predictions, which we validate empirically using 280K sentence pairs in English and Chinese: (1) correlation between the probability of strings within minimal pairs, i.e., string pairs with minimal semantic differences; (2) correlation between models' and humans' deltas within minimal pairs; and (3) poor separation in probability space between unpaired grammatical and ungrammatical strings. Our analyses give theoretical grounding for using probability to learn about LMs' structural knowledge, and suggest directions for future work in LM grammatical evaluation.

cs.CL↗

arXiv2025

BabyBabelLM: A Multilingual Benchmark of Developmentally Plausible Training Data

We present BabyBabelLM, a multilingual collection of datasets modeling the language a person observes from birth until they acquire a native language. We curate developmentally plausible pretraining data aiming to cover the equivalent of 100M English words of content in each of 45 languages. We compile evaluation suites and train baseline models in each language. BabyBabelLM aims to facilitate multilingual pretraining and cognitive modeling.

cs.CL↗

arXiv2025

Mix-modal Federated Learning for MRI Image Segmentation

Magnetic resonance imaging (MRI) image segmentation is crucial in diagnosing and treating many diseases, such as brain tumors. Existing MRI image segmentation methods mainly fall into a centralized multimodal paradigm, which is inapplicable in engineering non-centralized mix-modal medical scenarios. In this situation, each distributed client (hospital) processes multiple mixed MRI modalities, and the modality set and image data for each client are diverse, suffering from extensive client-wise modality heterogeneity and data heterogeneity. In this paper, we first formulate non-centralized mix-modal MRI image segmentation as a new paradigm for federated learning (FL) that involves multiple modalities, called mix-modal federated learning (MixMFL). It distinguishes from existing multimodal federating learning (MulMFL) and cross-modal federating learning (CroMFL) paradigms. Then, we proposed a novel modality decoupling and memorizing mix-modal federated learning framework (MDM-MixMFL) for MRI image segmentation, which is characterized by a modality decoupling strategy and a modality memorizing mechanism. Specifically, the modality decoupling strategy disentangles each modality into modality-tailored and modality-shared information. During mix-modal federated updating, corresponding modality encoders undergo tailored and shared updating, respectively. It facilitates stable and adaptive federating aggregation of heterogeneous data and modalities from distributed clients. Besides, the modality memorizing mechanism stores client-shared modality prototypes dynamically refreshed from every modality-tailored encoder to compensate for incomplete modalities in each local client.

cs.CV↗

arXiv2025

Privileged Self-Access Matters for Introspection in AI

Whether AI models can introspect is an increasingly important practical question. But there is no consensus on how introspection is to be defined. Beginning from a recently proposed ''lightweight'' definition, we argue instead for a thicker one. According to our proposal, introspection in AI is any process which yields information about internal states through a process more reliable than one with equal or lower computational cost available to a third party. Using experiments where LLMs reason about their internal temperature parameters, we show they can appear to have lightweight introspection while failing to meaningfully introspect per our proposed definition.

cs.AI↗

arXiv2025

Building Safer Sites: A Large-Scale Multi-Level Dataset for Construction Safety Research

Construction safety research is a critical field in civil engineering, aiming to mitigate risks and prevent injuries through the analysis of site conditions and human factors. However, the limited volume and lack of diversity in existing construction safety datasets pose significant challenges to conducting in-depth analyses. To address this research gap, this paper introduces the Construction Safety Dataset (CSDataset), a well-organized comprehensive multi-level dataset that encompasses incidents, inspections, and violations recorded sourced from the Occupational Safety and Health Administration (OSHA). This dataset uniquely integrates structured attributes with unstructured narratives, facilitating a wide range of approaches driven by machine learning and large language models. We also conduct a preliminary approach benchmarking and various cross-level analyses using our dataset, offering insights to inform and enhance future efforts in construction safety. For example, we found that complaint-driven inspections were associated with a 17.3% reduction in the likelihood of subsequent incidents. Our dataset and code are released at https://github.com/zhenhuiou/Construction-Safety-Dataset-CSDataset.

cs.LG↗

arXiv2025

Federated Client-tailored Adapter for Medical Image Segmentation

Medical image segmentation in X-ray images is beneficial for computer-aided diagnosis and lesion localization. Existing methods mainly fall into a centralized learning paradigm, which is inapplicable in the practical medical scenario that only has access to distributed data islands. Federated Learning has the potential to offer a distributed solution but struggles with heavy training instability due to client-wise domain heterogeneity (including distribution diversity and class imbalance). In this paper, we propose a novel Federated Client-tailored Adapter (FCA) framework for medical image segmentation, which achieves stable and client-tailored adaptive segmentation without sharing sensitive local data. Specifically, the federated adapter stirs universal knowledge in off-the-shelf medical foundation models to stabilize the federated training process. In addition, we develop two client-tailored federated updating strategies that adaptively decompose the adapter into common and individual components, then globally and independently update the parameter groups associated with common client-invariant and individual client-specific units, respectively. They further stabilize the heterogeneous federated learning process and realize optimal client-tailored instead of sub-optimal global-compromised segmentation models. Extensive experiments on three large-scale datasets demonstrate the effectiveness and superiority of the proposed FCA framework for federated medical segmentation.

cs.CV↗

arXiv2025

Language Models Fail to Introspect About Their Knowledge of Language

There has been recent interest in whether large language models (LLMs) can introspect about their own internal states. Such abilities would make LLMs more interpretable, and also validate the use of standard introspective methods in linguistics to evaluate grammatical knowledge in models (e.g., asking "Is this sentence grammatical?"). We systematically investigate emergent introspection across 21 open-source LLMs, in two domains where introspection is of theoretical interest: grammatical knowledge and word prediction. Crucially, in both domains, a model's internal linguistic knowledge can be theoretically grounded in direct measurements of string probability. We then evaluate whether models' responses to metalinguistic prompts faithfully reflect their internal knowledge. We propose a new measure of introspection: the degree to which a model's prompted responses predict its own string probabilities, beyond what would be predicted by another model with nearly identical internal knowledge. While both metalinguistic prompting and probability comparisons lead to high task accuracy, we do not find evidence that LLMs have privileged "self-access". By using general tasks, controlling for model similarity, and evaluating a wide range of open-source models, we show that LLMs cannot introspect, and add new evidence to the argument that prompted responses should not be conflated with models' linguistic generalizations.

cs.CL↗

arXiv2024

A Systematic Assessment of Language Models with Linguistic Minimal Pairs in Chinese

We present ZhoBLiMP, the largest linguistic minimal pair benchmark for Chinese, with over 100 paradigms, ranging from topicalization to the \textit{Ba} construction. We then train from scratch a suite of Chinese language models (LMs) with different tokenizers, parameter sizes, and token volumes, to study the learning curves of LMs on Chinese. To mitigate the biases introduced by unequal lengths of the sentences in a minimal pair, we propose a new metric named sub-linear length normalized log-probabilities (SLLN-LP). Using SLLN-LP as the metric, our results show that \textsc{Anaphor}, \textsc{Quantifiers}, and \textsc{Ellipsis} in Chinese are difficult for LMs even up to 32B parameters, and that SLLN-LP successfully mitigates biases in ZhoBLiMP, JBLiMP and BLiMP. We conclude that future evaluations should be more carefully designed to consider the intricate relations between linking functions, LMs, and targeted minimal pairs.

cs.CL↗