Continuous Self-Improvement of Large Language Models by Test-time Training with Verifier-Driven Sample Selection

Learning to adapt pretrained language models to unlabeled, out-of-distribution data is a critical challenge, as models often falter on structurally novel reasoning tasks even while excelling within their training distribution. We introduce a new framework called VDS-TTT - Verifier-Driven Sample Selection for Test-Time Training to efficiently address this. We use a learned verifier to score a pool of generated responses and select only from high ranking pseudo-labeled examples for fine-tuned adaptation. Specifically, for each input query our LLM generates N candidate answers; the verifier assigns a reliability score to each, and the response with the highest confidence and above a fixed threshold is paired with its query for test-time training. We fine-tune only low-rank LoRA adapter parameters, ensuring adaptation efficiency and fast convergence. Our proposed self-supervised framework is the first to synthesize verifier driven test-time training data for continuous self-improvement of the model. Experiments across three diverse benchmarks and three state-of-the-art LLMs demonstrate that VDS-TTT yields up to a 32.29% relative improvement over the base model and a 6.66% gain compared to verifier-based methods without test-time training, highlighting its effectiveness and efficiency for on-the-fly large language model adaptation.

Test-Time Training Done Right

Test-Time Training (TTT) models context dependencies by adapting part of the model's weights (referred to as fast weights) during inference. This fast weight, akin to recurrent states in RNNs, stores temporary memories of past tokens in the current sequence. Existing TTT methods struggled to show effectiveness in handling long-context data, due to their inefficiency on modern GPUs. The TTT layers in many of these approaches operate with extremely low FLOPs utilization (often <5%) because they deliberately apply small online minibatch sizes (e.g., updating fast weights every 16 or 64 tokens). Moreover, a small minibatch implies fine-grained block-wise causal dependencies in the data, unsuitable for data beyond 1D ordered sequences, like sets or N-dimensional grids such as images or videos. In contrast, we pursue the opposite direction by using an extremely large chunk update, ranging from 2K to 1M tokens across tasks of varying modalities, which we refer to as Large Chunk Test-Time Training (LaCT). It improves hardware utilization by orders of magnitude, and more importantly, facilitates scaling of nonlinear state size (up to 40% of model parameters), hence substantially improving state capacity, all without requiring cumbersome and error-prone kernel implementations. It also allows easy integration of sophisticated optimizers, e.g. Muon for online updates. We validate our approach across diverse modalities and tasks, including novel view synthesis with image set, language models, and auto-regressive video diffusion. Our approach can scale up to 14B-parameter AR video diffusion model on sequences up to 56K tokens. In our longest sequence experiment, we perform novel view synthesis with 1 million context length. We hope this work will inspire and accelerate new research in the field of long-context modeling and test-time training. Website: https://tianyuanzhang.com/projects/ttt-done-right

The MUSE Benchmark: Probing Music Perception and Auditory Relational Reasoning in Audio LLMS

Multimodal Large Language Models (MLLMs) have demonstrated capabilities in audio understanding, but current evaluations may obscure fundamental weaknesses in relational reasoning. We introduce the Music Understanding and Structural Evaluation (MUSE) Benchmark, an open-source resource with 10 tasks designed to probe fundamental music perception skills. We evaluate four SOTA models (Gemini Pro and Flash, Qwen2.5-Omni, and Audio-Flamingo 3) against a large human baseline (N=200). Our results reveal a wide variance in SOTA capabilities and a persistent gap with human experts. While Gemini Pro succeeds on basic perception, Qwen and Audio Flamingo 3 perform at or near chance, exposing severe perceptual deficits. Furthermore, we find Chain-of-Thought (CoT) prompting provides inconsistent, often detrimental results. Our work provides a critical tool for evaluating invariant musical representations and driving development of more robust AI systems.

CLiVR: Conversational Learning System in Virtual Reality with AI-Powered Patients

Simulations constitute a fundamental component of medical and nursing education and traditionally employ standardized patients (SP) and high-fidelity manikins to develop clinical reasoning and communication skills. However, these methods require substantial resources, limiting accessibility and scalability. In this study, we introduce CLiVR, a Conversational Learning system in Virtual Reality that integrates large language models (LLMs), speech processing, and 3D avatars to simulate realistic doctor-patient interactions. Developed in Unity and deployed on the Meta Quest 3 platform, CLiVR enables trainees to engage in natural dialogue with virtual patients. Each simulation is dynamically generated from a syndrome-symptom database and enhanced with sentiment analysis to provide feedback on communication tone. Through an expert user study involving medical school faculty (n=13), we assessed usability, realism, and perceived educational impact. Results demonstrated strong user acceptance, high confidence in educational potential, and valuable feedback for improvement. CLiVR offers a scalable, immersive supplement to SP-based training.

Inferred global dense residue transition graphs from primary structure sequences enable protein interaction prediction via directed graph convolutional neural networks

Introduction Accurate prediction of protein-protein interactions (PPIs) is crucial for understanding cellular functions and advancing drug development. Existing in-silico methods use direct sequence embeddings from Protein Language Models (PLMs). Others use Graph Neural Networks (GNNs) for 3D protein structures. This study explores less computationally intensive alternatives. We introduce a novel framework for downstream PPI prediction through link prediction. Methods We introduce a two-stage graph representation learning framework, ProtGram-DirectGCN. First, we developed ProtGram. This approach models a protein's primary structure as a hierarchy of globally inferred n-gram graphs. In these graphs, residue transition probabilities define edge weights. Each edge connects a pair of residues in a directed graph. The probabilities are aggregated from a large corpus of sequences. Second, we propose DirectGCN, a custom directed graph convolutional neural network. This model features a unique convolutional layer. It processes information through separate path-specific transformations: incoming, outgoing, and undirected. A shared transformation is also applied. These paths are combined via a learnable gating mechanism. We apply DirectGCN to ProtGram graphs to learn residue-level embeddings. These embeddings are pooled via attention to generate protein-level embeddings for prediction. Results We first established the efficacy of DirectGCN on standard node classification benchmarks. Its performance matches established methods on general datasets. The model excels at complex, directed graphs with dense, heterophilic structures. When applied to PPI prediction, the full ProtGram-DirectGCN framework delivers robust predictive power. This strong performance holds even with limited training data.

A Multimodal Approach to Heritage Preservation in the Context of Climate Change

Cultural heritage sites face accelerating degradation due to climate change, yet tradi- tional monitoring relies on unimodal analysis (visual inspection or environmental sen- sors alone) that fails to capture the complex interplay between environmental stres- sors and material deterioration. We propose a lightweight multimodal architecture that fuses sensor data (temperature, humidity) with visual imagery to predict degradation severity at heritage sites. Our approach adapts PerceiverIO with two key innovations: (1) simplified encoders (64D latent space) that prevent overfitting on small datasets (n=37 training samples), and (2) Adaptive Barlow Twins loss that encourages modality complementarity rather than redundancy. On data from Strasbourg Cathedral, our model achieves 76.9% accu- racy, a 43% improvement over standard multimodal architectures (VisualBERT, Trans- former) and 25% over vanilla PerceiverIO. Ablation studies reveal that sensor-only achieves 61.5% while image-only reaches 46.2%, confirming successful multimodal synergy. A systematic hyperparameter study identifies an optimal moderate correlation target ({\tau} =0.3) that balances align- ment and complementarity, achieving 69.2% accuracy compared to other {\tau} values ({\tau} =0.1/0.5/0.7: 53.8%, {\tau} =0.9: 61.5%). This work demonstrates that architectural sim- plicity combined with contrastive regularization enables effective multimodal learning in data-scarce heritage monitoring contexts, providing a foundation for AI-driven con- servation decision support systems.

Extracting latent representations from X-ray spectra. Classification, regression, and accretion signatures of Chandra sources

The study of X-ray spectra is crucial to understanding the physical nature of astrophysical sources. Machine learning methods can extract compact and informative representations of data from large datasets. The Chandra Source Catalog (CSC) provides a rich archive of X-ray spectral data, which remains largely underexplored in this context. This work aims to develop a compact and physically meaningful representation of Chandra X-ray spectra using deep learning. To verify that the learned representation captures relevant information, we evaluate it through classification, regression, and interpretability analyses. We use a transformer-based autoencoder to compress X-ray spectra. The input spectra, drawn from the CSC, include only high-significance detections. Astrophysical source types and physical summary statistics are compiled from external catalogs. We evaluate the learned representation in terms of spectral reconstruction accuracy, clustering performance on 8 known astrophysical source classes, and correlation with physical quantities such as hardness ratios and hydrogen column density ($N_H$). The autoencoder accurately reconstructs spectra with 8 latent variables. Clustering in the latent space yields a balanced classification accuracy of $\sim$40% across the 8 source classes, increasing to $\sim$69% when restricted to AGNs and stellar-mass compact objects exclusively. Moreover, latent features correlate with non-linear combinations of spectral fluxes, suggesting that the compressed representation encodes physically relevant information. The proposed autoencoder-based pipeline is a powerful tool for the representation and interpretation of X-ray spectra, providing a compact latent space that supports both classification and the estimation of physical properties. This work demonstrates the potential of deep learning for spectral studies and uncovering new patterns in X-ray data.

SketchGuard: Scaling Byzantine-Robust Decentralized Federated Learning via Sketch-Based Screening

Decentralized Federated Learning (DFL) enables privacy-preserving collaborative training without centralized servers, but remains vulnerable to Byzantine attacks where malicious clients submit corrupted model updates. Existing Byzantine-robust DFL defenses rely on similarity-based neighbor screening that requires every client to exchange and compare complete high-dimensional model vectors with all neighbors in each training round, creating prohibitive communication and computational costs that prevent deployment at web scale. We propose SketchGuard, a general framework that decouples Byzantine filtering from model aggregation through sketch-based neighbor screening. SketchGuard compresses $d$-dimensional models to $k$-dimensional sketches ($k \ll d$) using Count Sketch for similarity comparisons, then selectively fetches full models only from accepted neighbors, reducing per-round communication complexity from $O(d|N_i|)$ to $O(k|N_i| + d|S_i|)$, where $|N_i|$ is the neighbor count and $|S_i| \le |N_i|$ is the accepted neighbor count. We establish rigorous convergence guarantees in both strongly convex and non-convex settings, proving that Count Sketch compression preserves Byzantine resilience with controlled degradation bounds where approximation errors introduce only a $(1+O(\epsilon))$ factor in the effective threshold parameter. Comprehensive experiments across multiple datasets, network topologies, and attack scenarios demonstrate that SketchGuard maintains identical robustness to state-of-the-art methods while reducing computation time by up to 82% and communication overhead by 50-70% depending on filtering effectiveness, with benefits scaling multiplicatively with model dimensionality and network connectivity. These results establish the viability of sketch-based compression as a fundamental enabler of robust DFL at web scale.

A New Method of Constructing Hadamard Matrices, Circulant Hadamard Matrices, CZCS, GCS, CCC, and CZCSS

A Hadamard matrix $H$ is a square matrix of order $n$ with entries $\pm 1$, such that $HH^\top=nI_{n}$, where $I_n$ is an identity matrix of order $n$. A circulant Hadamard matrix $H$ is a Hadamard matrix that has rows of entries in cyclic order. There exist only $8$ circulant Hadamard matrices of order 4, and here, we provide a novel construction of all such $8$ circulant Hadamard matrices using a linear operator and generalized Boolean function (GBF). The constructed circulant Hadamard matrices are used recursively to construct a binary cross Z-complementary set (CZCS) of all lengths with an even phase, a binary Golay complementary set (GCS) of all lengths, and Hadamard matrices of order $2^{n+2}$, where $n\geq1$. The construction of a binary CZCS covering all lengths was not available before. We also propose an alternative, lower-complexity construction of binary GCSs of all lengths and Hadamard matrices of order $2^{a+1}10^b26^c$ using circulant matrices, where $ a,b,c \geq 0$. The proposed binary GCS covers all lengths with a flexible flock size. The constructions of GCS are further extended to form binary complete complementary code (CCC) of the parameter $(2N,2N,2N)-CCC$ where $N=2^a10^b26^c, a,b,c \geq 0$. The constructed binary CCC provides a flexible flock size. The construction of CZCS is further extended to form a binary optimal cross-Z complementary sequence set (CZCSS) of the parameter $(2^{n+2}, 2^{n+2}, 2^{n+2}, 2^{n+1})-CZCSS$, where $n\geq1$. Finally, we provide a relation between Hadamard matrices and GCS, which enables the study of the Hadamard conjecture in a new direction. We also provided a few properties of circulant matrices over aperiodic cross-correlation (ACCF) and aperiodic auto-correlation (AACF), which are used to prove the theorems. All proposed constructions are novel, and their parameters are compared with the existing state-of-the-art.

BoN Appetit Team at LeWiDi-2025: Best-of-N Test-time Scaling Can Not Stomach Annotation Disagreements (Yet)

Test-time scaling is a family of techniques to improve LLM outputs at inference time by performing extra computation. To the best of our knowledge, test-time scaling has been limited to domains with verifiably correct answers, like mathematics and coding. We transfer test-time scaling to the LeWiDi-2025 tasks to evaluate annotation disagreements. We experiment with three test-time scaling methods: two benchmark algorithms (Model Averaging and Majority Voting), and a Best-of-N sampling method. The two benchmark methods improve LLM performance consistently on the LeWiDi tasks, but the Best-of-N method does not. Our experiments suggest that the Best-of-N method does not currently transfer from mathematics to LeWiDi tasks, and we analyze potential reasons for this gap.