Compressing Transformer inputs into compressd tokens allows running LLMs with improved speed and cost efficiency. Based on the compression method ICAE, we carefully examine the position identifier choices for compressed tokens and also propose a new compression loss. We demonstrate empirically that our proposed methods achieve significantly higher compression ratios (15x compared to 4x for ICAE), while being able to attain comparable reconstruction performance.

To mitigate this selection bias problem, previous solutions utilized debiasing methods to adjust the model's input and/or output. Our work, in contrast, investigates the model's internal representation of the selection bias. Specifically, we introduce a novel debiasing approach, Bias Node Pruning (BNP), which eliminates the linear layer parameters that contribute to the bias. Furthermore, we present Auxiliary Option Injection (AOI), a simple yet effective input modification technique for debiasing, which is compatible even with black-box LLMs. To provide a more systematic evaluation of selection bias, we review existing metrics and introduce Choice Kullback-Leibler Divergence (CKLD), which addresses the insensitivity of the commonly used metrics to imbalance in choice labels. Experiments show that our methods are robust and adaptable across various datasets when applied to three LLMs. The advent of large language models (LLMs) has revolutionized artificial intelligence applications, particularly in the domain of natural language processing. These models have demonstrated outstanding performance across a variety of use cases, including chatbots, machine translation, text generation, data annotation, etc. Their ability to answer questions with high precision has opened up new avenues for automated systems. Despite their remarkable abilities, LLMs suffer from the selection bias problem that often occurs in answering multiplechoice questions (MCQs). When selecting the answer for an MCQ, many LLMs prefer the choices in a given position (e.g., the last choice), or with a specific choice symbol (e.g., (A) or (3)) (Zheng et al., 2024; Wei et al., 2024; Pezeshkpour & Hruschka, 2024). Many previous works have attempted to explain this phenomenon and/or propose diverse ways to mitigate selection bias. While there are a few works focused on either modifying the input format (Li et al., 2023b; Robinson et al., 2023) or calibrating the output probabilities (Zheng et al., 2024; Reif Figure 1: We propose BNP and & Schwartz, 2024; Wei et al., 2024), to the best of our knowledge, AOI to reduce selection bias for no embedding or parameter-level investigation has been white-box and black-box models. Because selection bias originates from internal The CKLD metric is also proposed parameter-level computations, it is crucial to explore how the to encourage a more standardized LLM embeddings contribute to the bias in their output responses. Understanding the internal representation of selection bias can help us combat it. By scrutinizing the interaction between the internal representation and the LLM parameters, we develop a novel approach to debias the model. Specifically, we propose Bias Node Pruning (BNP), which eliminates nodes in the final linear layer that contribute to selection bias. By dropping as few as 32 out of 4096 nodes in the final layer, we can significantly reduce selection bias and improve question-answering performance.

AI-driven program repair uses AI models to repair buggy software by producing patches. Rapid advancements in AI surely impact state-of-the-art performance of program repair. Yet, grasping this progress requires frequent and standardized evaluations. We propose RepairBench, a novel leaderboard for AI-driven program repair. The key characteristics of RepairBench are: 1) it is execution-based: all patches are compiled and executed against a test suite, 2) it assesses frontier models in a frequent and standardized way. RepairBench leverages two high-quality benchmarks, Defects4J and GitBug-Java, to evaluate frontier models against real-world program repair tasks. We publicly release the evaluation framework of RepairBench. We will update the leaderboard as new frontier models are released.

Assessing the performance of systems to classify Multi-Party Conversations (MPC) is challenging due to the interconnection between linguistic and structural characteristics of conversations. Conventional evaluation methods often overlook variances in model behavior across different levels of structural complexity on interaction graphs. In this work, we propose a methodological pipeline to investigate model performance across specific structural attributes of conversations. As a proof of concept we focus on Response Selection and Addressee Recognition tasks, to diagnose model weaknesses. To this end, we extract representative diagnostic subdatasets with a fixed number of users and a good structural variety from a large and open corpus of online MPCs. We further frame our work in terms of data minimization, avoiding the use of original usernames to preserve privacy, and propose alternatives to using original text messages. Results show that response selection relies more on the textual content of conversations, while addressee recognition requires capturing their structural dimension. Using an LLM in a zero-shot setting, we further highlight how sensitivity to prompt variations is task-dependent.

The Superficial Alignment Hypothesis posits that almost all of a language model's abilities and knowledge are learned during pre-training, while post-training is about giving a model the right style and format. We re-examine these claims by empirically studying the scaling behavior of post-training with increasing finetuning examples and evaluating them using objective task-specific standardized benchmarks. Through experiments with the Llama-3, Mistral, and Llama-2 model families of multiple sizes, we observe that, similar to the pre-training scaling laws, post-training task performance scales as a power law against the number of finetuning examples. This power law relationship holds across a broad array of capabilities, including mathematical reasoning, coding, instruction following, and multihop-reasoning. In addition, for tasks like math and multihop reasoning, we observe that a handful of examples merely align the model stylistically but do not saturate performance on the benchmarks. Model performance is instead correlated with its reasoning ability and it improves significantly with more examples, illustrating the need for holistic evaluation programs leveraging objective benchmarks in addition to measurement of alignment to human preferences. We also observe that language models are not necessarily limited to using knowledge learned during pre-training. With appropriate post-training, a model's ability to integrate new knowledge greatly improves on downstream tasks like multihop question-answering. Taken together, these results shed new light on the Superficial Alignment Hypothesis, suggesting that it is, at best, an over-simplification.

Open-ended questions test a more thorough understanding than closed-ended questions and are often a preferred assessment method. However, open-ended questions are tedious to grade and subject to personal bias. Therefore, there have been efforts to speed up the grading process through automation. Short Answer Grading (SAG) systems aim to automatically score students' answers. Despite growth in SAG methods and capabilities, there exists no comprehensive short-answer grading benchmark across different subjects, grading scales, and distributions. Thus, it is hard to assess the capabilities of current automated grading methods in terms of their generalizability. In this preliminary work, we introduce the combined ASAG2024 benchmark to facilitate the comparison of automated grading systems. Combining seven commonly used short-answer grading datasets in a common structure and grading scale. For our benchmark, we evaluate a set of recent SAG methods, revealing that while LLM-based approaches reach new high scores, they still are far from reaching human performance. This opens up avenues for future research on human-machine SAG systems.

Originally proposed for handling time series data, Auto-regressive Decision Trees (ARDTs) have not yet been explored for language modeling. This paper delves into both the theoretical and practical applications of ARDTs in this new context. We theoretically demonstrate that ARDTs can compute complex functions, such as simulating automata, Turing machines, and sparse circuits, by leveraging "chain-of-thought" computations. Our analysis provides bounds on the size, depth, and computational efficiency of ARDTs, highlighting their surprising computational power. Empirically, we train ARDTs on simple language generation tasks, showing that they can learn to generate coherent and grammatically correct text on par with a smaller Transformer model. Additionally, we show that ARDTs can be used on top of transformer representations to solve complex reasoning tasks. This research reveals the unique computational abilities of ARDTs, aiming to broaden the architectural diversity in language model development.

We study how differences in persuasive language across Wikipedia articles, written in either English and Russian, can uncover each culture's distinct perspective on different subjects. We develop a large language model (LLM) powered system to identify instances of persuasive language in multilingual texts. Instead of directly prompting LLMs to detect persuasion, which is subjective and difficult, we propose to reframe the task to instead ask high-level questions (HLQs) which capture different persuasive aspects. Importantly, these HLQs are authored by LLMs themselves. LLMs over-generate a large set of HLQs, which are subsequently filtered to a small set aligned with human labels for the original task. We then apply our approach to a large-scale, bilingual dataset of Wikipedia articles (88K total), using a two-stage identify-then-extract prompting strategy to find instances of persuasion. We quantify the amount of persuasion per article, and explore the differences in persuasion through several experiments on the paired articles. Notably, we generate rankings of articles by persuasion in both languages. These rankings match our intuitions on the culturally-salient subjects; Russian Wikipedia highlights subjects on Ukraine, while English Wikipedia highlights the Middle East. Grouping subjects into larger topics, we find politically-related events contain more persuasion than others. We further demonstrate that HLQs obtain similar performance when posed in either English or Russian. Our methodology enables cross-lingual, cross-cultural understanding at scale, and we release our code, prompts, and data.

Large Language Models (LLMs) have shown remarkable capabilities, but their inherent probabilistic nature often leads to inconsistency and inaccuracy in complex problem-solving tasks. This paper introduces DANA (Domain-Aware Neurosymbolic Agent), an architecture that addresses these issues by integrating domain-specific knowledge with neurosymbolic approaches. We begin by analyzing current AI architectures, including AutoGPT, LangChain ReAct and OpenAI's ChatGPT, through a neurosymbolic lens, highlighting how their reliance on probabilistic inference contributes to inconsistent outputs. In response, DANA captures and applies domain expertise in both natural-language and symbolic forms, enabling more deterministic and reliable problem-solving behaviors. We implement a variant of DANA using Hierarchical Task Plans (HTPs) in the open-source OpenSSA framework. This implementation achieves over 90\% accuracy on the FinanceBench financial-analysis benchmark, significantly outperforming current LLM-based systems in both consistency and accuracy. Application of DANA in physical industries such as semiconductor shows that its flexible architecture for incorporating knowledge is effective in mitigating the probabilistic limitations of LLMs and has potential in tackling complex, real-world problems that require reliability and precision.

Attention mechanisms, particularly softmax attention, have been instrumental in the success of transformer-based models such as GPT. However, the quadratic memory complexity of softmax attention with respect to sequence length poses significant challenges for processing longer sequences. We introduce Cottention, a novel attention mechanism that replaces the softmax operation with cosine similarity. By leveraging the properties of cosine similarity and rearranging the attention equation, Cottention achieves native linear memory complexity with respect to sequence length, making it inherently more memory-efficient than softmax attention. We demonstrate that Cottention can be reformulated as a recurrent neural network (RNN) with a finite hidden state, allowing for constant memory usage during inference. We evaluate Cottention on both the bidirectional BERT and causal GPT tasks, demonstrating comparable performance to softmax attention while significantly reducing memory requirements. To ensure efficient computation, we develop a custom CUDA kernel for Cottention. Our results show that Cottention is a promising alternative to softmax attention, enabling the processing of longer sequences without sacrificing performance, due to its native linear memory complexity and ability to maintain a constant memory footprint during inference.