Unveiling Language Skills under Circuits

The exploration of language skills in language models (LMs) has always been one of the central goals in mechanistic interpretability. However, existing circuit analyses often fall short in representing the full functional scope of these models, primarily due to the exclusion of Feed-Forward layers. Additionally, isolating the effect of a single language skill from a text, which inherently involves multiple entangled skills, poses a significant challenge. To address these gaps, we introduce a novel concept, Memory Circuit, a minimum unit that fully and independently manipulates the memory-reading functionality of a language model, and disentangle the transformer model precisely into a circuit graph which is an ensemble of paths connecting different memory circuits. Based on this disentanglement, we identify salient circuit paths, named as skill paths, responsible for three crucial language skills, i.e., the Previous Token Skill, Induction Skill and In-Context Learning (ICL) Skill, leveraging causal effect estimation through interventions and counterfactuals. Our experiments on various datasets confirm the correspondence between our identified skill paths and language skills, and validate three longstanding hypotheses: 1) Language skills are identifiable through circuit dissection; 2) Simple language skills reside in shallow layers, whereas complex language skills are found in deeper layers; 3) Complex language skills are formed on top of simpler language skills. Our codes are available at: https://github.com/Zodiark-ch/Language-Skill-of-LLMs.

Unraveling the Mechanics of Learning-Based Demonstration Selection for In-Context Learning

Large Language Models (LLMs) have demonstrated impressive in-context learning (ICL) capabilities from few-shot demonstration exemplars. While recent learning-based demonstration selection methods have proven beneficial to ICL by choosing more useful exemplars, their underlying mechanisms are opaque, hindering efforts to address limitations such as high training costs and poor generalization across tasks. These methods generally assume the selection process captures similarities between the exemplar and the target instance, however, it remains unknown what kinds of similarities are captured and vital to performing ICL. To dive into this question, we analyze the working mechanisms of the learning-based demonstration selection methods and empirically identify two important factors related to similarity measurement: 1) The ability to integrate different levels of task-agnostic text similarities between the input of exemplars and test cases enhances generalization power across different tasks. 2) Incorporating task-specific labels when measuring the similarities significantly improves the performance on each specific task. We validate these two findings through extensive quantitative and qualitative analyses across ten datasets and various LLMs. Based on our findings, we introduce two effective yet simplified exemplar selection methods catering to task-agnostic and task-specific demands, eliminating the costly LLM inference overhead.

Quantifying Emergence in Large Language Models

Emergence, broadly conceptualized as the ``intelligent'' behaviors of LLMs, has recently been studied and proved challenging to quantify due to the lack of a measurable definition. Most commonly, it has been estimated statistically through model performances across extensive datasets and tasks, which consumes significant resources. In addition, such estimation is difficult to interpret and may not accurately reflect the models' intrinsic emergence. In this work, we propose a quantifiable solution for estimating emergence. Inspired by emergentism in dynamics, we quantify the strength of emergence by comparing the entropy reduction of the macroscopic (semantic) level with that of the microscopic (token) level, both of which are derived from the representations within the transformer block. Using a low-cost estimator, our quantification method demonstrates consistent behaviors across a suite of LMs (GPT-2, GEMMA, etc.) under both in-context learning (ICL) and natural sentences. Empirical results show that (1) our method gives consistent measurements which align with existing observations based on performance metrics, validating the effectiveness of our emergence quantification; (2) our proposed metric uncovers novel emergence patterns such as the correlations between the variance of our metric and the number of ``shots'' in ICL, which further suggests a new way of interpreting hallucinations in LLMs; (3) we offer a potential solution towards estimating the emergence of larger and closed-resource LMs via smaller LMs like GPT-2. Our codes are available at: https://github.com/Zodiark-ch/Emergence-of-LLMs/.

Decomposing Label Space, Format and Discrimination: Rethinking How LLMs Respond and Solve Tasks via In-Context Learning

In-context Learning (ICL) has emerged as a powerful capability alongside the development of scaled-up large language models (LLMs). By instructing LLMs using few-shot demonstrative examples, ICL enables them to perform a wide range of tasks without updating millions of parameters. However, the precise contributions of demonstrations towards improving end-task performance have not been thoroughly investigated in recent analytical studies. In this paper, we empirically decompose the overall performance of ICL into three dimensions, label space, format, and discrimination, and we evaluate four general-purpose LLMs across a diverse range of tasks. Counter-intuitively, we find that the demonstrations have a marginal impact on provoking discriminative knowledge of language models. However, ICL exhibits significant efficacy in regulating the label space and format which helps LLMs to respond in desired label words. We then demonstrate this ability functions similar to detailed instructions for LLMs to follow. We additionally provide an in-depth analysis of the mechanism of retrieval helping with ICL and find that retrieving the most semantically similar examples notably boosts model's discriminative capability.

Re2LLM: Reflective Reinforcement Large Language Model for Session-based Recommendation

Large Language Models (LLMs) are emerging as promising approaches to enhance session-based recommendation (SBR), where both prompt-based and fine-tuning-based methods have been widely investigated to align LLMs with SBR. However, the former methods struggle with optimal prompts to elicit the correct reasoning of LLMs due to the lack of task-specific feedback, leading to unsatisfactory recommendations. Although the latter methods attempt to fine-tune LLMs with domain-specific knowledge, they face limitations such as high computational costs and reliance on open-source backbones. To address such issues, we propose a Reflective Reinforcement Large Language Model (Re2LLM) for SBR, guiding LLMs to focus on specialized knowledge essential for more accurate recommendations effectively and efficiently. In particular, we first design the Reflective Exploration Module to effectively extract knowledge that is readily understandable and digestible by LLMs. To be specific, we direct LLMs to examine recommendation errors through self-reflection and construct a knowledge base (KB) comprising hints capable of rectifying these errors. To efficiently elicit the correct reasoning of LLMs, we further devise the Reinforcement Utilization Module to train a lightweight retrieval agent. It learns to select hints from the constructed KB based on the task-specific feedback, where the hints can serve as guidance to help correct LLMs reasoning for better recommendations. Extensive experiments on multiple real-world datasets demonstrate that our method consistently outperforms state-of-the-art methods.

Can Language Models Act as Knowledge Bases at Scale?

Large language models (LLMs) have demonstrated remarkable proficiency in understanding and generating responses to complex queries through large-scale pre-training. However, the efficacy of these models in memorizing and reasoning among large-scale structured knowledge, especially world knowledge that explicitly covers abundant factual information remains questionable. Addressing this gap, our research investigates whether LLMs can effectively store, recall, and reason with knowledge on a large scale comparable to latest knowledge bases (KBs) such as Wikidata. Specifically, we focus on three crucial aspects to study the viability: (1) the efficiency of LLMs with different sizes in memorizing the exact knowledge in the large-scale KB; (2) the flexibility of recalling the memorized knowledge in response to natural language queries; (3) the capability to infer new knowledge through reasoning. Our findings indicate that while LLMs hold promise as large-scale KBs capable of retrieving and responding with flexibility, enhancements in their reasoning capabilities are necessary to fully realize their potential.

Backdoor Attacks on Dense Passage Retrievers for Disseminating Misinformation

Dense retrievers and retrieval-augmented language models have been widely used in various NLP applications. Despite being designed to deliver reliable and secure outcomes, the vulnerability of retrievers to potential attacks remains unclear, raising concerns about their security. In this paper, we introduce a novel scenario where the attackers aim to covertly disseminate targeted misinformation, such as hate speech or advertisement, through a retrieval system. To achieve this, we propose a perilous backdoor attack triggered by grammar errors in dense passage retrieval. Our approach ensures that attacked models can function normally for standard queries but are manipulated to return passages specified by the attacker when users unintentionally make grammatical mistakes in their queries. Extensive experiments demonstrate the effectiveness and stealthiness of our proposed attack method. When a user query is error-free, our model consistently retrieves accurate information while effectively filtering out misinformation from the top-k results. However, when a query contains grammar errors, our system shows a significantly higher success rate in fetching the targeted content.

TELLER: A Trustworthy Framework for Explainable, Generalizable and Controllable Fake News Detection

The proliferation of fake news has emerged as a severe societal problem, raising significant interest from industry and academia. While existing deep-learning based methods have made progress in detecting fake news accurately, their reliability may be compromised caused by the non-transparent reasoning processes, poor generalization abilities and inherent risks of integration with large language models (LLMs). To address this challenge, we propose {\methodname}, a novel framework for trustworthy fake news detection that prioritizes explainability, generalizability and controllability of models. This is achieved via a dual-system framework that integrates cognition and decision systems, adhering to the principles above. The cognition system harnesses human expertise to generate logical predicates, which guide LLMs in generating human-readable logic atoms. Meanwhile, the decision system deduces generalizable logic rules to aggregate these atoms, enabling the identification of the truthfulness of the input news across diverse domains and enhancing transparency in the decision-making process. Finally, we present comprehensive evaluation results on four datasets, demonstrating the feasibility and trustworthiness of our proposed framework. Our implementation is available at \url{https://github.com/less-and-less-bugs/Trust_TELLER}.

Minds versus Machines: Rethinking Entailment Verification with Language Models

Humans make numerous inferences in text comprehension to understand discourse. This paper aims to understand the commonalities and disparities in the inference judgments between humans and state-of-the-art Large Language Models (LLMs). Leveraging a comprehensively curated entailment verification benchmark, we evaluate both human and LLM performance across various reasoning categories. Our benchmark includes datasets from three categories (NLI, contextual QA, and rationales) that include multi-sentence premises and different knowledge types, thereby evaluating the inference capabilities in complex reasoning instances. Notably, our findings reveal LLMs' superiority in multi-hop reasoning across extended contexts, while humans excel in tasks necessitating simple deductive reasoning. Leveraging these insights, we introduce a fine-tuned Flan-T5 model that outperforms GPT-3.5 and rivals with GPT-4, offering a robust open-source solution for entailment verification. As a practical application, we showcase the efficacy of our finetuned model in enhancing self-consistency in model-generated explanations, resulting in a 6% performance boost on average across three multiple-choice question-answering datasets.

Training Language Models to Generate Text with Citations via Fine-grained Rewards

While recent Large Language Models (LLMs) have proven useful in answering user queries, they are prone to hallucination, and their responses often lack credibility due to missing references to reliable sources. An intuitive solution to these issues would be to include in-text citations referring to external documents as evidence. While previous works have directly prompted LLMs to generate in-text citations, their performances are far from satisfactory, especially when it comes to smaller LLMs. In this work, we propose an effective training framework using fine-grained rewards to teach LLMs to generate highly supportive and relevant citations, while ensuring the correctness of their responses. We also conduct a systematic analysis of applying these fine-grained rewards to common LLM training strategies, demonstrating its advantage over conventional practices. We conduct extensive experiments on Question Answering (QA) datasets taken from the ALCE benchmark and validate the model's generalizability using EXPERTQA. On LLaMA-2-7B, the incorporation of fine-grained rewards achieves the best performance among the baselines, even surpassing that of GPT-3.5-turbo.