Yizhou Sun

Hongzhe Du, Weikai Li, Min Cai et al.

Post-training is essential for the success of large language models (LLMs), transforming pre-trained base models into more useful and aligned post-trained models. While plenty of works have studied post-training algorithms and evaluated post-training models by their outputs, it remains understudied how post-training reshapes LLMs internally. In this paper, we compare base and post-trained LLMs mechanistically from four perspectives to better understand post-training effects. Our findings across model families and datasets reveal that: (1) Post-training does not change the factual knowledge storage locations, and it adapts knowledge representations from the base model while developing new knowledge representations; (2) Both truthfulness and refusal can be represented by vectors in the hidden representation space. The truthfulness direction is highly similar between the base and post-trained model, and it is effectively transferable for interventions; (3) The refusal direction is different between the base and post-trained models, and it shows limited forward transferability; (4) Differences in confidence between the base and post-trained models cannot be attributed to entropy neurons. Our study provides insights into the fundamental mechanisms preserved and altered during post-training, facilitates downstream tasks like model steering, and could potentially benefit future research in interpretability and LLM post-training. Our code is publicly available at https://github.com/HZD01/post-training-mechanistic-analysis.

Arjun Subramonian, Vagrant Gautam, Preethi Seshadri et al. · meta-ai

Numerous methods have been proposed to measure LLM misgendering, including probability-based evaluations (e.g., automatically with templatic sentences) and generation-based evaluations (e.g., with automatic heuristics or human validation). However, it has gone unexamined whether these evaluation methods have convergent validity, that is, whether their results align. Therefore, we conduct a systematic meta-evaluation of these methods across three existing datasets for LLM misgendering. We propose a method to transform each dataset to enable parallel probability- and generation-based evaluation. Then, by automatically evaluating a suite of 6 models from 3 families, we find that these methods can disagree with each other at the instance, dataset, and model levels, conflicting on 20.2% of evaluation instances. Finally, with a human evaluation of 2400 LLM generations, we show that misgendering behaviour is complex and goes far beyond pronouns, which automatic evaluations are not currently designed to capture, suggesting essential disagreement with human evaluations. Based on our findings, we provide recommendations for future evaluations of LLM misgendering. Our results are also more widely relevant, as they call into question broader methodological conventions in LLM evaluation, which often assume that different evaluation methods agree.

Guancheng Wan, Lucheng Fu, Haoxin Liu et al.

The performance of Large Language Models (LLMs) hinges on carefully engineered prompts. However, prevailing prompt optimization methods, ranging from heuristic edits and reinforcement learning to evolutionary search, primarily target point-wise accuracy. They seldom enforce paraphrase invariance or searching stability, and therefore cannot remedy this brittleness in practice. Automated prompt search remains brittle: small, semantically preserving paraphrases often cause large performance swings. We identify this brittleness as the textual sharpness of the prompt landscape. In this work, we provide the first formal treatment of textual sharpness in the discrete, semantic space of prompts, together with an operational robustness criterion over a semantic neighborhood; the design is black-box or API-only, requiring no gradients to update the model's parameters. Then we introduce TARE (Textual Sharpness-Aware Evolving), a derivative-free framework that alternates between an inner, sampling-based adversarial search that stresses a prompt with hard paraphrases and an outer, robust selection that prefers candidates whose neighborhoods remain strong. We further propose ATARE, which learns anisotropic weights to shape the semantic neighborhood and adapts its radius over time to balance exploration and fidelity. Diverse tasks evaluate our methods, whose design for minimizing textual sharpness gap leads to prompts that preserve accuracy under paraphrasing, outperforming accuracy-only prompt search while remaining computationally practical.

John Harvill, Ziwei Fan, Hao Wang et al.

Existing work on prompt compression for Large Language Models (LLM) focuses on lossy methods that try to maximize the retention of semantic information that is relevant to downstream tasks while significantly reducing the sequence length. In this paper, we introduce a task-agnostic lossless compression technique similar to LZ77 that makes it possible to reduce the input token sequence length on average by 27\% and 18\% for the two evaluation tasks explored here. Given that we use transformer-based LLMs, this equates to 47\% and 33\% less encoding computation, respectively, due to the quadratic nature of attention. The token sequence transformation is trivial to reverse and highlights that no semantic information is lost in the process. We evaluate our proposed approach on two tasks that require strict preservation of semantics/syntax and demonstrate that existing lossy compression methods perform poorly in this setting. We find that our lossless compression technique produces only a small gap in performance compared to using the uncompressed input and posit that larger models and an expanded computing budget would likely erase the gap entirely.

Zimin Zhang, Qianli Wu, Botao Xia et al.

Artificial intelligence (AI) is transforming scientific research, with explainable AI methods like concept-based models (CMs) showing promise for new discoveries. However, in molecular science, CMs are less common than black-box models like Graph Neural Networks (GNNs), due to their need for predefined concepts and manual labeling. This paper introduces the Automated Molecular Concept (AutoMolCo) framework, which leverages Large Language Models (LLMs) to automatically generate and label predictive molecular concepts. Through iterative concept refinement, AutoMolCo enables simple linear models to outperform GNNs and LLM in-context learning on several benchmarks. The framework operates without human knowledge input, overcoming limitations of existing CMs while maintaining explainability and allowing easy intervention. Experiments on MoleculeNet and High-Throughput Experimentation (HTE) datasets demonstrate that AutoMolCo-induced explainable CMs are beneficial for molecular science research.

Shichang Zhang, Da Zheng, Jiani Zhang et al.

Text-rich graphs, prevalent in data mining contexts like e-commerce and academic graphs, consist of nodes with textual features linked by various relations. Traditional graph machine learning models, such as Graph Neural Networks (GNNs), excel in encoding the graph structural information, but have limited capability in handling rich text on graph nodes. Large Language Models (LLMs), noted for their superior text understanding abilities, offer a solution for processing the text in graphs but face integration challenges due to their limitation for encoding graph structures and their computational complexities when dealing with extensive text in large neighborhoods of interconnected nodes. This paper introduces ``Hierarchical Compression'' (HiCom), a novel method to align the capabilities of LLMs with the structure of text-rich graphs. HiCom processes text in a node's neighborhood in a structured manner by organizing the extensive textual information into a more manageable hierarchy and compressing node text step by step. Therefore, HiCom not only preserves the contextual richness of the text but also addresses the computational challenges of LLMs, which presents an advancement in integrating the text processing power of LLMs with the structural complexities of text-rich graphs. Empirical results show that HiCom can outperform both GNNs and LLM backbones for node classification on e-commerce and citation graphs. HiCom is especially effective for nodes from a dense region in a graph, where it achieves a 3.48% average performance improvement on five datasets while being more efficient than LLM backbones.