Recent large language models (LLMs) have excelled across a wide range of tasks, but their use in high-stakes and compute-limited settings has intensified the demand for interpretability and efficiency. We address this need by proposing Induction-head ngram models (Induction-Gram), a method that builds an efficient, interpretable LM by bolstering modern ngram models with a hand-engineered "induction head". This induction head uses a custom neural similarity metric to efficiently search the model's input context for potential next-word completions. This process enables Induction-Gram to provide ngram-level grounding for each generated token. Moreover, experiments show that this simple method significantly improves next-word prediction over baseline interpretable models (up to 26%p) and can be used to speed up LLM inference for large models through speculative decoding. We further study Induction-Gram in a natural-language neuroscience setting, where the goal is to predict the next fMRI response in a sequence. It again provides a significant improvement over interpretable models (20% relative increase in the correlation of predicted fMRI responses), potentially enabling deeper scientific investigation of language selectivity in the brain. The code is available at https://github.com/ejkim47/induction-gram.

Accurate document retrieval is crucial for the success of retrieval-augmented generation (RAG) applications, including open-domain question answering and code completion. While large language models (LLMs) have been employed as dense encoders or listwise rerankers in RAG systems, they often struggle with reasoning-intensive tasks because they lack nuanced analysis when judging document relevance. To address this limitation, we introduce JudgeRank, a novel agentic reranker that emulates human cognitive processes when assessing document relevance. Our approach consists of three key steps: (1) query analysis to identify the core problem, (2) document analysis to extract a query-aware summary, and (3) relevance judgment to provide a concise assessment of document relevance. We evaluate JudgeRank on the reasoning-intensive BRIGHT benchmark, demonstrating substantial performance improvements over first-stage retrieval methods and outperforming other popular reranking approaches. In addition, JudgeRank performs on par with fine-tuned state-of-the-art rerankers on the popular BEIR benchmark, validating its zero-shot generalization capability. Through comprehensive ablation studies, we demonstrate that JudgeRank's performance generalizes well across LLMs of various sizes while ensembling them yields even more accurate reranking than individual models.

Open-weight large language model (LLM) zoos allow users to quickly integrate state-of-the-art models into systems. Despite increasing availability, selecting the most appropriate model for a given task still largely relies on public benchmark leaderboards and educated guesses. This can be unsatisfactory for both inference service providers and end users, where the providers usually prioritize cost efficiency, while the end users usually prioritize model output quality for their inference requests. In commercial settings, these two priorities are often brought together in Service Level Agreements (SLA). We present MESS+, an online stochastic optimization algorithm for energy-optimal model selection from a model zoo, which works on a per-inference-request basis. For a given SLA that requires high accuracy, we are up to 2.5x more energy efficient with MESS+ than with randomly selecting an LLM from the zoo while maintaining SLA quality constraints.

Vision Language Models (VLMs) have been successful at many chart comprehension tasks that require attending to both the images of charts and their accompanying textual descriptions. However, it is not well established how VLM performance profiles map to human-like behaviors. If VLMs can be shown to have human-like chart comprehension abilities, they can then be applied to a broader range of tasks, such as designing and evaluating visualizations for human readers. This paper lays the foundations for such applications by evaluating the accuracy of zero-shot prompting of VLMs on graphical perception tasks with established human performance profiles. Our findings reveal that VLMs perform similarly to humans under specific task and style combinations, suggesting that they have the potential to be used for modeling human performance. Additionally, variations to the input stimuli show that VLM accuracy is sensitive to stylistic changes such as fill color and chart contiguity, even when the underlying data and data mappings are the same.

The increasing frequency of suicidal thoughts highlights the importance of early detection and intervention. Social media platforms, where users often share personal experiences and seek help, could be utilized to identify individuals at risk. However, the large volume of daily posts makes manual review impractical. This paper explores the use of Large Language Models (LLMs) to automatically detect suicidal content in text-based social media posts. We propose a novel method for generating pseudo-labels for unlabeled data by prompting LLMs, along with traditional classification fine-tuning techniques to enhance label accuracy. To create a strong suicide detection model, we develop an ensemble approach involving prompting with Qwen2-72B-Instruct, and using fine-tuned models such as Llama3-8B, Llama3.1-8B, and Gemma2-9B. We evaluate our approach on the dataset of the Suicide Ideation Detection on Social Media Challenge, a track of the IEEE Big Data 2024 Big Data Cup. Additionally, we conduct a comprehensive analysis to assess the impact of different models and fine-tuning strategies on detection performance. Experimental results show that the ensemble model significantly improves the detection accuracy, by 5% points compared with the individual models. It achieves a weight F1 score of 0.770 on the public test set, and 0.731 on the private test set, providing a promising solution for identifying suicidal content in social media. Our analysis shows that the choice of LLMs affects the prompting performance, with larger models providing better accuracy. Our code and checkpoints are publicly available at https://github.com/khanhvynguyen/Suicide_Detection_LLMs.

Recently, there has been an explosion of large language models created through fine-tuning with data from larger models. These small models able to produce outputs that appear qualitatively similar to significantly larger models. However, one of the key limitations that have been observed with these models is their propensity to hallucinate significantly more often than larger models. In particular, they have been observed to generate coherent outputs that involve factually incorrect information and spread misinformation, toxicity, and stereotypes. There are many potential causes of hallucination, of which, one hypothesis is that fine-tuning a model on data produced by a larger model leads to a knowledge mismatch which contributes to hallucination. In particular, it is hypothesized that there is a mismatch between the knowledge that is fed to the model to fine-tune it and the knowledge that is already present in the graph. Fine-tuning the model on data that has such mismatch could contribute to an increased propensity to hallucinate. We show that on an unseen test set, a smaller model fine-tuned on data generated from a larger model produced more wrong answers when compared to models fine-tuned on data created by the small model, which confirms the hypothesis.

In real-world applications, the distribution of the data, and our goals, evolve over time. The prevailing theoretical framework for studying machine learning, namely probably approximately correct (PAC) learning, largely ignores time. As a consequence, existing strategies to address the dynamic nature of data and goals exhibit poor real-world performance. This paper develops a theoretical framework called "Prospective Learning" that is tailored for situations when the optimal hypothesis changes over time. In PAC learning, empirical risk minimization (ERM) is known to be consistent. We develop a learner called Prospective ERM, which returns a sequence of predictors that make predictions on future data. We prove that the risk of prospective ERM converges to the Bayes risk under certain assumptions on the stochastic process generating the data. Prospective ERM, roughly speaking, incorporates time as an input in addition to the data. We show that standard ERM as done in PAC learning, without incorporating time, can result in failure to learn when distributions are dynamic. Numerical experiments illustrate that prospective ERM can learn synthetic and visual recognition problems constructed from MNIST and CIFAR-10.

Large language models (LLM) hold significant potential for applications in biomedicine, but they struggle with hallucinations and outdated knowledge. While retrieval-augmented generation (RAG) is generally employed to address these issues, it also has its own set of challenges: (1) LLMs are vulnerable to irrelevant or incorrect context, (2) medical queries are often not well-targeted for helpful information, and (3) retrievers are prone to bias toward the specific source corpus they were trained on. In this study, we present RAG$^2$ (RAtionale-Guided RAG), a new framework for enhancing the reliability of RAG in biomedical contexts. RAG$^2$ incorporates three key innovations: a small filtering model trained on perplexity-based labels of rationales, which selectively augments informative snippets of documents while filtering out distractors; LLM-generated rationales as queries to improve the utility of retrieved snippets; a structure designed to retrieve snippets evenly from a comprehensive set of four biomedical corpora, effectively mitigating retriever bias. Our experiments demonstrate that RAG$^2$ improves the state-of-the-art LLMs of varying sizes, with improvements of up to 6.1\%, and it outperforms the previous best medical RAG model by up to 5.6\% across three medical question-answering benchmarks. Our code is available at https://github.com/dmis-lab/RAG2.

When examined through the lens of their residual streams, a puzzling property emerges in transformer networks: residual contributions (e.g., attention heads) sometimes specialize in specific tasks or input attributes. In this paper, we analyze this phenomenon in vision transformers, focusing on the spectral geometry of residuals, and explore its implications for modality alignment in vision-language models. First, we link it to the intrinsically low-dimensional structure of visual head representations, zooming into their principal components and showing that they encode specialized roles across a wide variety of input data distributions. Then, we analyze the effect of head specialization in multimodal models, focusing on how improved alignment between text and specialized heads impacts zero-shot classification performance. This specialization-performance link consistently holds across diverse pre-training data, network sizes, and objectives, demonstrating a powerful new mechanism for boosting zero-shot classification through targeted alignment. Ultimately, we translate these insights into actionable terms by introducing ResiDual, a technique for spectral alignment of the residual stream. Much like panning for gold, it lets the noise from irrelevant unit principal components (i.e., attributes) wash away to amplify task-relevant ones. Remarkably, this dual perspective on modality alignment yields fine-tuning level performances on different data distributions while modeling an extremely interpretable and parameter-efficient transformation, as we extensively show on more than 50 (pre-trained network, dataset) pairs.

The large models, as predicted by scaling raw forecasts, have made groundbreaking progress in many fields, particularly in natural language generation tasks, where they have approached or even surpassed human levels. However, the unprecedented scale of their parameters brings significant computational and storage costs. These large models require substantial computational resources and GPU memory to operate. When adapting large models to specific downstream tasks, their massive parameter scale poses a significant challenge in fine-tuning on hardware platforms with limited computational power and GPU memory. To address this issue, Parameter-Efficient Fine-Tuning (PEFT) offers a practical solution by efficiently adjusting the parameters of large pre-trained models to suit various downstream tasks. Specifically, PEFT adjusts the parameters of pre-trained large models to adapt to specific tasks or domains, minimizing the introduction of additional parameters and the computational resources required. This review mainly introduces the preliminary knowledge of PEFT, the core ideas and principles of various PEFT algorithms, the applications of PEFT, and potential future research directions. By reading this review, we believe that interested parties can quickly grasp the PEFT methodology, thereby accelerating its development and innovation.