Deep neural networks excel in high-dimensional problems, outperforming models such as kernel methods, which suffer from the curse of dimensionality. However, the theoretical foundations of this success remain poorly understood. We follow the idea that the compositional structure of the learning task is the key factor determining when deep networks outperform other approaches. Taking a step towards formalizing this idea, we consider a simple compositional model, namely the multi-index model (MIM). In this context, we introduce and study hyper-kernel ridge regression (HKRR), an approach blending neural networks and kernel methods. Our main contribution is a sample complexity result demonstrating that HKRR can adaptively learn MIM, overcoming the curse of dimensionality. Further, we exploit the kernel nature of the estimator to develop ad hoc optimization approaches. Indeed, we contrast alternating minimization and alternating gradient methods both theoretically and numerically. These numerical results complement and reinforce our theoretical findings.

Data hiding is essential for secure communication across digital media, and recent advances in Deep Neural Networks (DNNs) provide enhanced methods for embedding secret information effectively. However, previous audio hiding methods often result in unsatisfactory quality when recovering secret audio, due to their inherent limitations in the modeling of time-frequency relationships. In this paper, we explore these limitations and introduce a new DNN-based approach. We use a flow-based invertible neural network to establish a direct link between stego audio, cover audio, and secret audio, enhancing the reversibility of embedding and extracting messages. To address common issues from time-frequency transformations that degrade secret audio quality during recovery, we implement a time-frequency loss on the time-domain signal. This approach not only retains the benefits of time-frequency constraints but also enhances the reversibility of message recovery, which is vital for practical applications. We also add an encryption technique to protect the hidden data from unauthorized access. Experimental results on the VCTK and LibriSpeech datasets demonstrate that our method outperforms previous approaches in terms of subjective and objective metrics and exhibits robustness to various types of noise, suggesting its utility in targeted secure communication scenarios.

Reinforcement learning from human or AI feedback (RLHF/RLAIF) has become the standard paradigm for aligning large language models (LLMs). However, most pipelines rely on a single reward model (RM), limiting alignment quality and risking overfitting. Recent work explores RM routing--dynamically selecting an RM from a candidate pool to exploit complementary strengths while maintaining O(1) RM calls--but existing methods suffer from cold-start and insufficient exploration. We propose BayesianRouter, a hybrid routing framework that combines offline RM strengths learning with online Bayesian selection. In the offline stage, a multi-task router is trained on preference data to estimate per-RM reliability. In the online stage, a Bayesian Thompson sampling router performs per-query RM selection, initializing RM-specific weight vectors with offline embeddings as Gaussian priors and adaptively updating their posteriors with online rewards to adapt to the evolving policy distribution. Extensive experiments on instruction-following (AlpacaEval-2, Arena-Hard, MT -Bench) and reasoning (GSM8K, MMLU) benchmarks show that BayesianRouter consistently outperforms individual RMs, RM ensembling, and existing routing methods. Large language models (LLMs) have revolutionized artificial intelligence, demonstrating substantial capabilities in language understanding, reasoning, and open-ended text generation across diverse domains (Guo et al., 2025; Achiam et al., 2023). To safely and effectively deploy LLMs, recent post-training techniques, particularly reinforcement learning from human feedback (RLHF) and its AI-augmented variant (RLAIF) (Wang et al., 2024b), aimed at aligning LLMs with human values and preferences. In standard RLHF, a reward model (RM) provides the feedback signal for optimizing the policy LLM; for example, in PPO, the RM provides a scalar reward to directly increase the probability of preferred responses (Ziegler et al., 2019), while in direct preference optimization (DPO), the RM compares two candidate responses to determine which is better (Dong et al., 2024).

Recent progress in retrieval-augmented generation (RAG) has led to more accurate and interpretable multi-hop question answering (QA). Yet, challenges persist in integrating iterative reasoning steps with external knowledge retrieval. To address this, we introduce StepChain GraphRAG, a framework that unites question decomposition with a Breadth-First Search (BFS) Reasoning Flow for enhanced multi-hop QA. Our approach first builds a global index over the corpus; at inference time, only retrieved passages are parsed on-the-fly into a knowledge graph, and the complex query is split into sub-questions. For each sub-question, a BFS-based traversal dynamically expands along relevant edges, assembling explicit evidence chains without overwhelming the language model with superfluous context. Experiments on MuSiQue, 2WikiMultiHopQA, and HotpotQA show that StepChain GraphRAG achieves state-of-the-art Exact Match and F1 scores. StepChain GraphRAG lifts average EM by 2.57% and F1 by 2.13% over the SOTA method, achieving the largest gain on HotpotQA (+4.70% EM, +3.44% F1). StepChain GraphRAG also fosters enhanced explainability by preserving the chain-of-thought across intermediate retrieval steps. We conclude by discussing how future work can mitigate the computational overhead and address potential hallucinations from large language models to refine efficiency and reliability in multi-hop QA.

Neural linguistic steganography aims to embed information into natural text while preserving statistical undetectability. A fundamental challenge in this ffeld stems from tokenization ambiguity in modern tokenizers, which can lead to catastrophic decoding failures. The recent method, SyncPool, addresses this ambiguity by employing a coarse-grained synchronization mechanism over groups of ambiguous candidates. However, SyncPool sacriffces embedding capacity, as it utilizes the entire Shannon entropy of an ambiguous group solely for synchronization rather than for payload embedding. We propose a method named look-ahead Sync, which overcomes the capacity limitation of SyncPool while retaining its provable security guarantees. Our approach performs minimal synchronized sampling only on truly indistinguishable token sequences, while strategically preserving all other discernible paths to maximize embedding capacity. We provide theoretical proofs for the security of our method and analyze the gap between its achievable embedding capacity and the theoretical upper bound. Experiments on English (using Llama 3) and Chinese (using Qwen 2.5) benchmarks show that our method consistently approaches the theoretical capacity upper bound and signiffcantly outperforms SyncPool. The improvement in embedding rate exceeds 160% in English and 25% in Chinese, particularly in settings with larger candidate pools. This work represents a signiffcant step toward practical high-capacity provably secure linguistic steganography.

While language models (LMs) offer great potential for conversational recommender systems (CRSs), the paucity of public CRS data makes fine-tuning LMs for CRSs challenging. In response, LMs as user simulators qua data generators can be used to train LM-based CRSs, but often lack behavioral consistency, generating utterance sequences inconsistent with those of any real user. To address this, we develop a methodology for generating natural dialogues that are consistent with a user's underlying state using behavior simulators together with LM-prompting. We illustrate our approach by generating a large, open-source CRS data set with both preference elicitation and example critiquing. Rater evaluation on some of these dialogues shows them to exhibit considerable consistency, factuality and naturalness.

Training long-context language models to capture long-range dependencies requires specialized data construction. Current approaches, such as generic text concatenation or heuristic-based variants, frequently fail to guarantee genuine long-range dependencies. We propose EntropyLong, a novel data construction method that leverages predictive uncertainty to verify dependency quality. Our approach identifies high-entropy positions in documents, retrieves semantically relevant contexts from large corpora, and verifies their utility by assessing whether they reduce prediction entropy. This model-in-the-loop verification ensures each dependency represents measurable information gain rather than spurious correlation. We construct training samples with long-range dependencies by combining original documents with these verified contextual supplements. Using FineWebEdu and Cosmopedia, we generate a dataset of 128K-length sequences with verified dependencies. Models trained on this data demonstrate significant improvements on RULER benchmarks, particularly in tasks requiring distant information. Following instruction fine-tuning, our models also achieve substantial gains on LongBenchv2, demonstrating enhanced long-context understanding. Extensive ablation studies further validate the necessity and effectiveness of entropybased verification for long-context training.

Empirical scaling laws establish a relationship between the number of parameters and model capability, as evidenced by models with hundreds of billions of parameters achieving state-of-the-art results on benchmarks (Kaplan et al., 2020; Grattafiori et al., 2024). However, the autoregressive generation process requires accessing all model parameters for each forward pass, creating a memory bandwidth bottleneck that impacts token generation latency. Furthermore, current trends toward more sophisticated LLM applications, such as multi-hop reasoning (Wei et al., 2022), tool integration (Patil et al., 2024), and reasoning capability (Y ang et al., 2025; DeepMind, 2025), produce longer output sequences, amplifying the computational burden of inference. One prominent approach to address inference latency is Speculative Decoding (SD), which achieves partial parallelization of the generation process (Leviathan et al., 2023; Chen et al., 2023). Standard SD operates by deploying a computationally efficient draft model to propose candidate token sequences, which are subsequently validated in parallel by the target model (the model of interest). The acceptance criterion for draft tokens relies on a probability-based alignment verification: draft tokens are accepted when their likelihood under the target model meets or exceeds their likelihood under the draft model.

Medical Multimodal Large Language Models (Med-MLLMs) have shown great promise in medical visual question answering (Med-VQA). However, when deployed in low-resource settings where abundant labeled data are unavailable, existing Med-MLLMs commonly fail due to their medical reasoning capability bottlenecks: (i) the intrinsic reasoning bottleneck that ignores the details from the medical image; (ii) the extrinsic reasoning bottleneck that fails to incorporate specialized medical knowledge. To address those limitations, we propose AMANDA, a training-free agentic framework that performs medical knowledge augmentation via LLM agents. Specifically, our intrinsic medical knowledge augmentation focuses on coarse-to-fine question decomposition for comprehensive diagnosis, while extrinsic medical knowledge augmentation grounds the reasoning process via biomedical knowledge graph retrieval. Extensive experiments across eight Med-VQA benchmarks demonstrate substantial improvements in both zero-shot and few-shot Med-VQA settings. The code is available at https://github.com/REAL-Lab-NU/AMANDA.

Large language models accelerate literature synthesis but can hallucinate and mis-cite, limiting their usefulness in expert workflows. We present RA-FSM (Research Assistant - Finite State Machine), a modular GPT-based research assistant that wraps generation in a finite-state control loop: Relevance -> Confidence -> Knowledge. The system is grounded in vector retrieval and a deterministic citation pipeline. The controller filters out-of-scope queries, scores answerability, decomposes questions, and triggers retrieval only when needed, and emits answers with confidence labels and in-corpus, de-duplicated references. A ranked-tier ingestion workflow constructs a domain knowledge base from journals, conferences, indices, preprints, and patents, writing both to a dense vector index and to a relational store of normalized metrics. We implement the system for photonics and evaluate it on six task categories: analytical reasoning, numerical analysis, methodological critique, comparative synthesis, factual extraction, and application design. In blinded A/B reviews, domain experts prefer RA-FSM to both a strong Notebook LM (NLM) and a vanilla Default GPT API call single-pass baseline, citing stronger boundary-condition handling and more defensible evidence use. Coverage and novelty analyses indicate that RA-FSM explores beyond the NLM while incurring tunable latency and cost overheads. The design emphasizes transparent, well-cited answers for high-stakes technical work and is generalizable to other scientific domains.