In this paper we explore where information is collected and how it is propagated throughout layers in large language models (LLMs). We begin by examining the surprising computational importance of punctuation tokens which previous work has identified as attention sinks and memory aids. Using intervention-based techniques, we evaluate the necessity and sufficiency (for preserving model performance) of punctuation tokens across layers in GPT-2, DeepSeek, and Gemma. Our results show stark model-specific differences: for GPT-2, punctuation is both necessary and sufficient in multiple layers, while this holds far less in DeepSeek and not at all in Gemma. Extending beyond punctuation, we ask whether LLMs process different components of input (e.g., subjects, adjectives, punctuation, full sentences) by forming early static summaries reused across the network, or if the model remains sensitive to changes in these components across layers. Extending beyond punctuation, we investigate whether different reasoning rules are processed differently by LLMs. In particular, through interchange intervention and layer-swapping experiments, we find that conditional statements (if, then), and universal quantification (for all) are processed very differently. Our findings offer new insight into the internal mechanisms of punctuation usage and reasoning in LLMs and have implications for interpretability.

Large language models (LLMs) have shown promise in medical domains, but their ability to handle specialized neurological reasoning requires systematic evaluation. We developed a comprehensive benchmark using 305 questions from Israeli Board Certification Exams in Neurology, classified along three complexity dimensions: factual knowledge depth, clinical concept integration, and reasoning complexity. We evaluated ten LLMs using base models, retrieval-augmented generation (RAG), and a novel multi-agent system. Results showed significant performance variation. OpenAI-o1 achieved the highest base performance (90.9% accuracy), while specialized medical models performed poorly (52.9% for Meditron-70B). RAG provided modest benefits but limited effectiveness on complex reasoning questions. In contrast, our multi-agent framework, decomposing neurological reasoning into specialized cognitive functions including question analysis, knowledge retrieval, answer synthesis, and validation, achieved dramatic improvements, especially for mid-range models. The LLaMA 3.3-70B-based agentic system reached 89.2% accuracy versus 69.5% for its base model, with substantial gains on level 3 complexity questions. The multi-agent approach transformed inconsistent subspecialty performance into uniform excellence, addressing neurological reasoning challenges that persisted with RAG enhancement. We validated our approach using an independent dataset of 155 neurological cases from MedQA. Results confirm that structured multi-agent approaches designed to emulate specialized cognitive processes significantly enhance complex medical reasoning, offering promising directions for AI assistance in challenging clinical contexts.

Large Language Models (LLMs) are increasingly deployed across diverse applications that demand balancing multiple, often conflicting, objectives -- such as helpfulness, harmlessness, or humor. Aligning outputs to user-specific preferences in such multi-objective settings typically requires fine-tuning models for each objective or preference configuration, which is computationally expensive and inflexible. We introduce MAVIS -- Multi-Objective Alignment via Value-Guided Inference-Time Search -- a lightweight inference-time alignment framework that enables dynamic control over LLM behavior without modifying the base model's weights. MAVIS trains a set of small value models, each corresponding to a distinct objective. At inference time, these value models are combined using user-specified weights to produce a tilting function that adjusts the base model's output distribution toward desired trade-offs. The value models are trained using a simple iterative algorithm that ensures monotonic improvement of the KL-regularized policy. We show empirically that MAVIS outperforms baselines that fine-tune per-objective models and combine them post hoc, and even approaches the performance of the idealized setting where models are fine-tuned for a user's exact preferences.

Recent advances in large language model (LLM) pretraining have shown that simply scaling data quantity eventually leads to diminishing returns, hitting a data wall. In response, the use of synthetic data for pretraining has emerged as a promising paradigm for pushing the frontier of performance. Despite this, the factors affecting synthetic data quality remain poorly understood. In this work, we introduce BeyondWeb, a synthetic data generation framework that produces high-quality synthetic data for pretraining. BeyondWeb significantly extends the capabilities of traditional web-scale datasets, outperforming state-of-the-art synthetic pretraining datasets such as Cosmopedia and Nemotron-CC's high-quality synthetic subset (Nemotron-Synth) by up to 5.1 percentage points (pp) and 2.6pp, respectively, when averaged across a suite of 14 benchmark evaluations. It delivers up to 7.7x faster training than open web data and 2.7x faster than Nemotron-Synth. Remarkably, a 3B model trained for 180B tokens on BeyondWeb outperforms an 8B model trained for the same token budget on Cosmopedia. We also present several insights from BeyondWeb on synthetic data for pretraining: what drives its benefits, which data to rephrase and how, and the impact of model size and family on data quality. Overall, our work shows that there's no silver bullet for generating high-quality synthetic pretraining data. The best outcomes require jointly optimizing many factors, a challenging task that requires rigorous science and practical expertise. Naive approaches can yield modest improvements, potentially at great cost, while well-executed methods can yield transformative improvements, as exemplified by BeyondWeb.

We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: https://github.com/JungHeeKim29/calibration-token.

Biodiversity assessments are critically affected by the spatial scale at which species richness is measured. How species richness accumulates with sampling area depends on natural and anthropogenic processes whose effects can change depending on the spatial scale considered. These accumulation dynamics, described by the species-area relationship (SAR), are challenging to assess because most biodiversity surveys are restricted to sampling areas much smaller than the scales at which these processes operate. Here, we combine sampling theory and deep learning to predict local species richness within arbitrarily large sampling areas, enabling for the first time to estimate spatial differences in SARs. We demonstrate our approach by predicting vascular plant species richness across Europe and evaluate predictions against an independent dataset of plant community inventories. The resulting model, named deep SAR, delivers multi-scale species richness maps, improving coarse grain richness estimates by 32% compared to conventional methods, while delivering finer grain estimates. Additional to its predictive capabilities, we show how our deep SAR model can provide fundamental insights on the multi-scale effects of key biodiversity processes. The capacity of our approach to deliver comprehensive species richness estimates across the full spectrum of ecologically relevant scales is essential for robust biodiversity assessments and forecasts under global change.

Generative AI (GenAI) is expected to play a pivotal role in enabling autonomous optimization in future wireless networks. Within the ORAN architecture, Large Language Models (LLMs) can be specialized to generate xApps and rApps by leveraging specifications and API definitions from the RAN Intelligent Controller (RIC) platform. However, fine-tuning base LLMs for telecom-specific tasks remains expensive and resource-intensive. Retrieval-Augmented Generation (RAG) offers a practical alternative through in-context learning, enabling domain adaptation without full retraining. While traditional RAG systems rely on vector-based retrieval, emerging variants such as GraphRAG and Hybrid GraphRAG incorporate knowledge graphs or dual retrieval strategies to support multi-hop reasoning and improve factual grounding. Despite their promise, these methods lack systematic, metric-driven evaluations, particularly in high-stakes domains such as ORAN. In this study, we conduct a comparative evaluation of Vector RAG, GraphRAG, and Hybrid GraphRAG using ORAN specifications. We assess performance across varying question complexities using established generation metrics: faithfulness, answer relevance, context relevance, and factual correctness. Results show that both GraphRAG and Hybrid GraphRAG outperform traditional RAG. Hybrid GraphRAG improves factual correctness by 8%, while GraphRAG improves context relevance by 11%.

Path recommendation (PR) aims to generate travel paths that are customized to a user's specific preferences and constraints. Conventional approaches often employ explicit optimization objectives or specialized machine learning architectures; however, these methods typically exhibit limited flexibility and generalizability, necessitating costly retraining to accommodate new scenarios. This paper introduces an alternative paradigm that conceptualizes PR as a natural language generation task. We present PathGPT, a retrieval-augmented large language model (LLM) system that leverages historical trajectory data and natural language user constraints to generate plausible paths. The proposed methodology first converts raw trajectory data into a human-interpretable textual format, which is then stored in a database. Subsequently, a hybrid retrieval system extracts path-specific context from this database to inform a pretrained LLM. The primary contribution of this work is a novel framework that demonstrates how integrating established information retrieval and generative model components can enable adaptive, zero-shot path generation across diverse scenarios. Extensive experiments on large-scale trajectory datasets indicate that PathGPT's performance is competitive with specialized, learning-based methods, underscoring its potential as a flexible and generalizable path generation system that avoids the need for retraining inherent in previous data-driven models.

This paper studies the problem of estimating physical properties (system identification) through visual observations.