Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for language model (LM) alignment. At its core, RLHF uses a margin-based loss for preference optimization, specifying ideal LM behavior only by the difference between preferred and dispreferred responses. In this paper, we identify a common pitfall of margin-based methods -- the under-specification of ideal LM behavior on preferred and dispreferred responses individually, which leads to two unintended consequences as the margin increases: (1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures. (2) The probability of preferred responses may decrease, even when those responses are ideal. We demystify the reasons behind these problematic behaviors: margin-based losses couple the change in the preferred probability to the gradient of the dispreferred one, and vice versa, often preventing the preferred probability from increasing while the dispreferred one decreases, and thus causing a synchronized increase or decrease in both probabilities. We term this effect, inherent in margin-based objectives, gradient entanglement. Formally, we derive conditions for general margin-based alignment objectives under which gradient entanglement becomes concerning: the inner product of the gradients of preferred and dispreferred log-probabilities is large relative to the individual gradient norms. We theoretically investigate why such inner products can be large when aligning language models and empirically validate our findings. Empirical implications of our framework extend to explaining important differences in the training dynamics of various preference optimization algorithms, and suggesting potential algorithm designs to mitigate the under-specification issue of margin-based methods and thereby improving language model alignment.

The recently introduced path-star task is a minimal task designed to exemplify limitations to the abilities of language models (Bachmann and Nagarajan, 2024). It involves a path-star graph where multiple arms radiate from a single starting node and each node is unique. Given the start node and a specified target node that ends an arm, the task is to generate the arm containing that target node. This is straightforward for a human but surprisingly difficult for language models, which did not outperform the random baseline. The authors hypothesized this is due to a deficiency in teacher-forcing and the next-token prediction paradigm. We demonstrate the task is learnable using teacher-forcing in alternative settings and that the issue is partially due to representation. We introduce a regularization method using structured samples of the same graph but with differing target nodes, improving results across a variety of model types. We provide RASP proofs showing the task is theoretically solvable. Finally, we find settings where an encoder-only model can consistently solve the task.

Recent work in machine learning community proposed multiple methods for performing lossy compression (quantization) of large matrices. This quantization is important for accelerating matrix multiplication (main component of large language models), which is often bottlenecked by the speed of loading these matrices from memory. Unlike classical vector quantization and rate-distortion theory, the goal of these new compression algorithms is to be able to approximate not the matrices themselves, but their matrix product. Specifically, given a pair of real matrices $A,B$ an encoder (compressor) is applied to each of them independently producing descriptions with $R$ bits per entry. These representations subsequently are used by the decoder to estimate matrix product $A^\top B$. In this work, we provide a non-asymptotic lower bound on the mean squared error of this approximation (as a function of rate $R$) for the case of matrices $A,B$ with iid Gaussian entries. Algorithmically, we construct a universal quantizer based on nested lattices with an explicit guarantee of approximation error for any (non-random) pair of matrices $A$, $B$ in terms of only Frobenius norms $\|A\|_F, \|B\|_F$ and $\|A^\top B\|_F$. For iid Gaussian matrices our quantizer achieves the lower bound and is, thus, asymptotically optimal. A practical low-complexity version of our quantizer achieves performance quite close to optimal. In information-theoretic terms we derive rate-distortion function for matrix multiplication of iid Gaussian matrices.

Scaling up autoregressive models in vision has not proven as beneficial as in large language models. In this work, we investigate this scaling problem in the context of text-to-image generation, focusing on two critical factors: whether models use discrete or continuous tokens, and whether tokens are generated in a random or fixed raster order using BERT- or GPT-like transformer architectures. Our empirical results show that, while all models scale effectively in terms of validation loss, their evaluation performance -- measured by FID, GenEval score, and visual quality -- follows different trends. Models based on continuous tokens achieve significantly better visual quality than those using discrete tokens. Furthermore, the generation order and attention mechanisms significantly affect the GenEval score: random-order models achieve notably better GenEval scores compared to raster-order models. Inspired by these findings, we train Fluid, a random-order autoregressive model on continuous tokens. Fluid 10.5B model achieves a new state-of-the-art zero-shot FID of 6.16 on MS-COCO 30K, and 0.69 overall score on the GenEval benchmark. We hope our findings and results will encourage future efforts to further bridge the scaling gap between vision and language models.

Watermarking techniques offer a promising way to identify machine-generated content via embedding covert information into the contents generated from language models (LMs). However, the robustness of the watermarking schemes has not been well explored. In this paper, we present De-mark, an advanced framework designed to remove n-gram-based watermarks effectively. Our method utilizes a novel querying strategy, termed random selection probing, which aids in assessing the strength of the watermark and identifying the red-green list within the n-gram watermark. Experiments on popular LMs, such as Llama3 and ChatGPT, demonstrate the efficiency and effectiveness of De-mark in watermark removal and exploitation tasks.

We present Legal Argument Reasoning (LAR), a novel task designed to evaluate the legal reasoning capabilities of Large Language Models (LLMs). The task requires selecting the correct next statement (from multiple choice options) in a chain of legal arguments from court proceedings, given the facts of the case. We constructed a dataset (LAR-ECHR) for this task using cases from the European Court of Human Rights (ECHR). We evaluated seven general-purpose LLMs on LAR-ECHR and found that (a) the ranking of the models is aligned with that of LegalBench, an established US-based legal reasoning benchmark, even though LAR-ECHR is based on EU law, (b) LAR-ECHR distinguishes top models more clearly, compared to LegalBench, (c) even the best model (GPT-4o) obtains 75.8% accuracy on LAR-ECHR, indicating significant potential for further model improvement. The process followed to construct LAR-ECHR can be replicated with cases from other legal systems.

While prior work has explored whether large language models (LLMs) possess a "theory of mind" (ToM) - the ability to attribute mental states to oneself and others - there has been little work testing whether LLMs can implicitly apply such knowledge to predict behavior, or to judge whether an observed behavior is rational. Such skills are critical for appropriate interaction in social environments. We create a new dataset, SimpleTom, containing concise, diverse stories (e.g., "The can of Pringles has moldy chips in it. Mary picks up the can in the supermarket and walks to the cashier."), each with three questions that test different degrees of ToM reasoning, asking models to predict (a) mental state ("Is Mary aware of the mold?"), (b) behavior ("Will Mary pay for the chips or report the mold?"), and (c) judgment ("Mary paid for the chips. Was that reasonable?"). To our knowledge, SimpleToM is the first dataset to systematically explore downstream reasoning requiring knowledge of mental states in realistic scenarios. Our experimental results are intriguing: While most models can reliably predict mental state on our dataset (a), they often fail to correctly predict the behavior (b), and fare even worse at judging whether given behaviors are reasonable (c), despite being correctly aware of the protagonist's mental state should make such secondary predictions obvious. We further show that we can help models do better at (b) and (c) via interventions such as reminding the model of its earlier mental state answer and mental-state-specific chain-of-thought prompting, raising the action prediction accuracies (e.g., from 49.5% to 93.5% for GPT-4o) and judgment accuracies (e.g., from 15.3% to 94.7% in GPT-4o). While this shows that models can be coaxed to perform well, it requires task-specific interventions, and the natural model performances remain low, a cautionary tale for LLM deployment.

Large Language Models (LLMs) have demonstrated remarkable performance across various domains, including healthcare. However, their ability to effectively represent structured non-textual data, such as the alphanumeric medical codes used in records like ICD-10 or SNOMED-CT, is limited and has been particularly exposed in recent research. This paper examines the challenges LLMs face in processing medical codes due to the shortcomings of current tokenization methods. As a result, we introduce the UniStruct architecture to design a multimodal medical foundation model of unstructured text and structured data, which addresses these challenges by adapting subword tokenization techniques specifically for the structured medical codes. Our approach is validated through model pre-training on both an extensive internal medical database and a public repository of structured medical records. Trained on over 1 billion tokens on the internal medical database, the proposed model achieves up to a 23% improvement in evaluation metrics, with around 2% gain attributed to our proposed tokenization. Additionally, when evaluated on the EHRSHOT public benchmark with a 1/1000 fraction of the pre-training data, the UniStruct model improves performance on over 42% of the downstream tasks. Our approach not only enhances the representation and generalization capabilities of patient-centric models but also bridges a critical gap in representation learning models' ability to handle complex structured medical data, alongside unstructured text.

Transformers have excelled in natural language processing and computer vision, paving their way to sensor-based Human Activity Recognition (HAR). Previous studies show that transformers outperform their counterparts exclusively when they harness abundant data or employ compute-intensive optimization algorithms. However, neither of these scenarios is viable in sensor-based HAR due to the scarcity of data in this field and the frequent need to perform training and inference on resource-constrained devices. Our extensive investigation into various implementations of transformer-based versus non-transformer-based HAR using wearable sensors, encompassing more than 500 experiments, corroborates these concerns. We observe that transformer-based solutions pose higher computational demands, consistently yield inferior performance, and experience significant performance degradation when quantized to accommodate resource-constrained devices. Additionally, transformers demonstrate lower robustness to adversarial attacks, posing a potential threat to user trust in HAR.

Large language models (LLMs) have recently gained much attention in building autonomous agents. However, the performance of current LLM-based web agents in long-horizon tasks is far from optimal, often yielding errors such as repeatedly buying a non-refundable flight ticket. By contrast, humans can avoid such an irreversible mistake, as we have an awareness of the potential outcomes (e.g., losing money) of our actions, also known as the "world model". Motivated by this, our study first starts with preliminary analyses, confirming the absence of world models in current LLMs (e.g., GPT-4o, Claude-3.5-Sonnet, etc.). Then, we present a World-model-augmented (WMA) web agent, which simulates the outcomes of its actions for better decision-making. To overcome the challenges in training LLMs as world models predicting next observations, such as repeated elements across observations and long HTML inputs, we propose a transition-focused observation abstraction, where the prediction objectives are free-form natural language descriptions exclusively highlighting important state differences between time steps. Experiments on WebArena and Mind2Web show that our world models improve agents' policy selection without training and demonstrate our agents' cost- and time-efficiency compared to recent tree-search-based agents.