Long-context large language models (LLMs) have achieved remarkable advancements, driven by techniques like Rotary Position Embedding (RoPE) (Su et al., 2023) and its extensions (Chen et al., 2023; Liu et al., 2024c; Peng et al., 2023). By adjusting RoPE parameters and incorporating training data with extended contexts, we can train performant models with considerably longer input sequences. However, existing RoPE-based methods exhibit performance limitations when applied to extended context lengths. This paper presents a comprehensive analysis of various attention mechanisms, including RoPE, No Positional Embedding (NoPE), and Query-Key Normalization (QK-Norm), identifying their strengths and shortcomings in long-context modeling. Our investigation identifies distinctive attention patterns in these methods and highlights their impact on long-context performance, providing valuable insights for architectural design. Building on these findings, we propose a novel architectural based on a hybrid attention mechanism that not only surpasses conventional RoPE-based transformer models in long context tasks but also achieves competitive performance on benchmarks requiring shorter context lengths.

Aligning Large Language Models (LLMs) with human values and away from undesirable behaviors (such as hallucination) has become increasingly important. Recently, steering LLMs towards a desired behavior via activation editing has emerged as an effective method to mitigate harmful generations at inference-time. Activation editing modifies LLM representations by preserving information from positive demonstrations (e.g., truthful) and minimising information from negative demonstrations (e.g., hallucinations). When these demonstrations come from a private dataset, the aligned LLM may leak private information contained in those private samples. In this work, we present the first study of aligning LLM behavior with private datasets. Our work proposes the Private Steering for LLM Alignment (PSA) algorithm to edit LLM activations with differential privacy (DP) guarantees. We conduct extensive experiments on seven different benchmarks with opensource LLMs of different sizes (0.5B to 7B) and model families (LlaMa, Qwen, Mistral and Gemma). Our results show that PSA achieves DP guarantees for LLM alignment with minimal loss in performance, including alignment metrics, openended text generation quality, and general-purpose reasoning. We also develop the first Membership Inference Attack (MIA) for evaluating and auditing the empirical privacy for the problem of LLM steering via activation editing. Our attack is tailored for activation editing and relies solely on the generated texts without their associated probabilities. Our experiments support the theoretical guarantees by showing improved guarantees for our PSA algorithm compared to several existing non-private techniques. LLMs often generate inaccurate, biased or even harmful information that violates human values and preferences (Rawte et al., 2023). In response, recent research has increasingly focused on aligning LLMs towards certain desired behaviors (Konen et al., 2024) while preventing potentially harmful and unsafe outcomes. This has led to the development of several techniques for aligning LLMs, such as Reinforcement Learning from Human Feedback (RLHF) (Ouyang et al., 2022), instruction tuning (Wei et al., 2022), In-Context Learning (ICL) (Dong et al., 2022), and prompt engineering (Cheng et al., 2024). Nevertheless, several challenges remain, including the lack of diverse and representative datasets for alignment (Liu et al., 2024c), difficulties in addressing out-of-distribution issues (Liu et al., 2024a), the choice of alignment strategy (Ivison et al., 2024) and the lack of interpretability in traditional alignment methods (Lee et al., 2024). The linear representation hypothesis (Park et al., 2024b) suggests that high-level concepts are linearly represented as directions in the representation space of LLMs.

Although many game-theoretic models replicate real interactions that often rely on natural language, explicit study of games where language is central to strategic interaction remains limited. This paper introduces the \emph{conversation game}, a multi-stage, extensive-form game based on linguistic strategic interaction. We focus on a subset of the games, called verdict games. In a verdict game, two players alternate to contribute to a conversation, which is evaluated at each stage by a non-strategic judge who may render a conclusive binary verdict, or a decision to continue the dialogue. The game ends once a limit is reached or a verdict is given. We show many familiar processes, such as interrogation or a court process fall under this category. We also, show that the Turing test is an instance of verdict game, and discuss the significance of a strategic view of the Turing test in the age of advanced AI deception. We show the practical relevance of the proposed concepts by simulation experiments, and show that a strategic agent outperforms a naive agent by a high margin.

We introduce DeltaLLM, a new post-training compression technique to reduce the memory footprint of LLMs. We propose an alternative way of structuring LLMs with weight sharing between layers in subsequent Transformer blocks, along with additional low-rank difference matrices between them. For training, we adopt the progressing module replacement method and show that the lightweight training of the low-rank modules with approximately 30M-40M tokens is sufficient to achieve performance on par with LLMs of comparable sizes trained from scratch. We release the resultant models, DeltaLLAMA and DeltaPHI, with a 12% parameter reduction, retaining 90% of the performance of the base Llama and Phi models on common knowledge and reasoning benchmarks. Our method also outperforms compression techniques JointDrop, LaCo, ShortGPT and SliceGPT with the same number of parameters removed. For example, DeltaPhi 2.9B with a 24% reduction achieves similar average zero-shot accuracies as recovery fine-tuned SlicedPhi 3.3B with a 12% reduction, despite being approximately 400M parameters smaller with no fine-tuning applied. This work provides new insights into LLM architecture design and compression methods when storage space is critical.

We study how a one-layer attention-only transformer develops relevant structures while learning to sort lists of numbers. At the end of training, the model organizes its attention heads in two main modes that we refer to as vocabulary-splitting and copy-suppression. Both represent simpler modes than having multiple heads handle overlapping ranges of numbers. Interestingly, vocabulary-splitting is present regardless of whether we use weight decay, a common regularization technique thought to drive simplification, supporting the thesis that neural networks naturally prefer simpler solutions. We relate copy-suppression to a mechanism in GPT-2 and investigate its functional role in our model. Guided by insights from a developmental analysis of the model, we identify features in the training data that drive the model's acquired final solution. This provides a concrete example of how the training data shape the internal organization of transformers, paving the way for future studies that could help us better understand how LLMs develop their internal structures.

We introduce the State Stream Transformer (SST), a novel LLM architecture that reveals emergent reasoning behaviours and capabilities latent in pretrained weights through addressing a fundamental limitation in traditional transformer models: the lack of latent computational continuity across autoregressive generations in the state space. SST introduces a sliding window latent state (FFN) cache with weighted decay that maintains and evolves persistent latent processes throughout autoregressive generations. Through controlled experiments comparing base and SST architectures using the same frozen weights, we demonstrate that this architectural modification alone enables enhanced reasoning capabilities which appear best explained by some form of potential higher-order processing, as evidenced by emergent metacognitive behaviours. These behaviours persist under controlled conditions designed to eliminate confounding factors such as stochastic variation or learned response patterns. Analysis of latent state distributions and processing dynamics provides evidence that it is solely the 'state stream' that is responsible for these phenomena. In quantitative evaluations, the SST achieves substantial performance improvements over the base model on two reasoning benchmarks, reaching 89.01\% accuracy on GSM-8K (0-shot) and 91.04\% on ARC Challenge (0-shot CoT). These findings indicate that persistent computation in the latent state space enables fundamentally different information processing and internal reasoning strategies, with implications for our understanding of artificial intelligence systems.

This paper presents an exploratory study that harnesses the capabilities of large language models (LLMs) to mine key ecological entities from invasion biology literature. Specifically, we focus on extracting species names, their locations, associated habitats, and ecosystems, information that is critical for understanding species spread, predicting future invasions, and informing conservation efforts. Traditional text mining approaches often struggle with the complexity of ecological terminology and the subtle linguistic patterns found in these texts. By applying general-purpose LLMs without domain-specific fine-tuning, we uncover both the promise and limitations of using these models for ecological entity extraction. In doing so, this study lays the groundwork for more advanced, automated knowledge extraction tools that can aid researchers and practitioners in understanding and managing biological invasions.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.

Guesstimation, the task of making approximate quantity estimates, is a common real-world challenge. However, it has been largely overlooked in large language models (LLMs) and vision language models (VLMs) research. We introduce a novel guesstimation dataset, MARBLES. This dataset requires one to estimate how many items (e.g., marbles) can fit into containers (e.g., a one-cup measuring cup), both with and without accompanying images. Inspired by the social science concept of the ``Wisdom of Crowds'' (WOC) - taking the median from estimates from a crowd), which has proven effective in guesstimation, we propose ``WOC decoding'' strategy for LLM guesstimation. We show that LLMs/VLMs perform well on guesstimation, suggesting that they possess some level of a "world model" necessary for guesstimation. Moreover, similar to human performance, the WOC decoding method improves LLM/VLM guesstimation accuracy. Furthermore, the inclusion of images in the multimodal condition enhances model performance. These results highlight the value of WOC decoding strategy for LLMs/VLMs and position guesstimation as a probe for evaluating LLMs/VLMs' world model. As LLMs' world model is a fundamental prerequisite for many real-world tasks, e.g., human-AI teaming, our findings have broad implications for the AI community.

The greatest ambition of mechanistic interpretability is to completely rewrite deep neural networks in a format that is more amenable to human understanding, while preserving their behavior and performance. In this paper, we attempt to partially rewrite a large language model using simple natural language explanations. We first approximate one of the feedforward networks in the LLM with a wider MLP with sparsely activating neurons - a transcoder - and use an automated interpretability pipeline to generate explanations for these neurons. We then replace the first layer of this sparse MLP with an LLM-based simulator, which predicts the activation of each neuron given its explanation and the surrounding context. Finally, we measure the degree to which these modifications distort the model's final output. With our pipeline, the model's increase in loss is statistically similar to entirely replacing the sparse MLP output with the zero vector. We employ the same protocol, this time using a sparse autoencoder, on the residual stream of the same layer and obtain similar results. These results suggest that more detailed explanations are needed to improve performance substantially above the zero ablation baseline.