CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation

Chain-of-Thought (CoT) enhances Large Language Models (LLMs) by enabling step-by-step reasoning in natural language. However, the language space may be suboptimal for reasoning. While implicit CoT methods attempt to enable reasoning without explicit CoT tokens, they have consistently lagged behind explicit CoT method in task performance. We propose CODI (Continuous Chain-of-Thought via Self-Distillation), a novel framework that distills CoT into a continuous space, where a shared model acts as both teacher and student, jointly learning explicit and implicit CoT while aligning their hidden activation on the token generating the final answer. CODI is the first implicit CoT method to match explicit CoT's performance on GSM8k while achieving 3.1x compression, surpassing the previous state-of-the-art by 28.2% in accuracy. Furthermore, CODI demonstrates scalability, robustness, and generalizability to more complex CoT datasets. Additionally, CODI retains interpretability by decoding its continuous thoughts, making its reasoning process transparent. Our findings establish implicit CoT as not only a more efficient but a powerful alternative to explicit CoT.

Two Heads Are Better Than One: Dual-Model Verbal Reflection at Inference-Time

Large Language Models (LLMs) often struggle with complex reasoning scenarios. While preference optimization methods enhance reasoning performance through training, they often lack transparency in why one reasoning outcome is preferred over another. Verbal reflection techniques improve explainability but are limited in LLMs' critique and refinement capacity. To address these challenges, we introduce a contrastive reflection synthesis pipeline that enhances the accuracy and depth of LLM-generated reflections. We further propose a dual-model reasoning framework within a verbal reinforcement learning paradigm, decoupling inference-time self-reflection into specialized, trained models for reasoning critique and refinement. Extensive experiments show that our framework outperforms traditional preference optimization methods across all evaluation metrics. Our findings also show that "two heads are better than one", demonstrating that a collaborative Reasoner-Critic model achieves superior reasoning performance and transparency, compared to single-model approaches.

RGAR: Recurrence Generation-augmented Retrieval for Factual-aware Medical Question Answering

Medical question answering requires extensive access to specialized conceptual knowledge. The current paradigm, Retrieval-Augmented Generation (RAG), acquires expertise medical knowledge through large-scale corpus retrieval and uses this knowledge to guide a general-purpose large language model (LLM) for generating answers. However, existing retrieval approaches often overlook the importance of factual knowledge, which limits the relevance of retrieved conceptual knowledge and restricts its applicability in real-world scenarios, such as clinical decision-making based on Electronic Health Records (EHRs). This paper introduces RGAR, a recurrence generation-augmented retrieval framework that retrieves both relevant factual and conceptual knowledge from dual sources (i.e., EHRs and the corpus), allowing them to interact and refine each another. Through extensive evaluation across three factual-aware medical question answering benchmarks, RGAR establishes a new state-of-the-art performance among medical RAG systems. Notably, the Llama-3.1-8B-Instruct model with RGAR surpasses the considerably larger, RAG-enhanced GPT-3.5. Our findings demonstrate the benefit of extracting factual knowledge for retrieval, which consistently yields improved generation quality.

RoleMRC: A Fine-Grained Composite Benchmark for Role-Playing and Instruction-Following

Role-playing is important for Large Language Models (LLMs) to follow diverse instructions while maintaining role identity and the role's pre-defined ability limits. Existing role-playing datasets mostly contribute to controlling role style and knowledge boundaries, but overlook role-playing in instruction-following scenarios. We introduce a fine-grained role-playing and instruction-following composite benchmark, named RoleMRC, including: (1) Multi-turn dialogues between ideal roles and humans, including free chats or discussions upon given passages; (2) Role-playing machine reading comprehension, involving response, refusal, and attempts according to passage answerability and role ability; (3) More complex scenarios with nested, multi-turn and prioritized instructions. The final RoleMRC features a 10.2k role profile meta-pool, 37.9k well-synthesized role-playing instructions, and 1.4k testing samples. We develop a pipeline to quantitatively evaluate the fine-grained role-playing and instruction-following capabilities of several mainstream LLMs, as well as models that are fine-tuned on our data. Moreover, cross-evaluation on external role-playing datasets confirms that models fine-tuned on RoleMRC enhances instruction-following without compromising general role-playing and reasoning capabilities. We also probe the neural-level activation maps of different capabilities over post-tuned LLMs. Access to our RoleMRC, RoleMRC-mix and Codes: https://github.com/LuJunru/RoleMRC.

WebWalker: Benchmarking LLMs in Web Traversal

Retrieval-augmented generation (RAG) demonstrates remarkable performance across tasks in open-domain question-answering. However, traditional search engines may retrieve shallow content, limiting the ability of LLMs to handle complex, multi-layered information. To address it, we introduce WebWalkerQA, a benchmark designed to assess the ability of LLMs to perform web traversal. It evaluates the capacity of LLMs to traverse a website's subpages to extract high-quality data systematically. We propose WebWalker, which is a multi-agent framework that mimics human-like web navigation through an explore-critic paradigm. Extensive experimental results show that WebWalkerQA is challenging and demonstrates the effectiveness of RAG combined with WebWalker, through the horizontal and vertical integration in real-world scenarios.

SCOPE: Optimizing Key-Value Cache Compression in Long-context Generation

Key-Value (KV) cache has become a bottleneck of LLMs for long-context generation. Despite the numerous efforts in this area, the optimization for the decoding phase is generally ignored. However, we believe such optimization is crucial, especially for long-output generation tasks based on the following two observations: (i) Excessive compression during the prefill phase, which requires specific full context impairs the comprehension of the reasoning task; (ii) Deviation of heavy hitters occurs in the reasoning tasks with long outputs. Therefore, SCOPE, a simple yet efficient framework that separately performs KV cache optimization during the prefill and decoding phases, is introduced. Specifically, the KV cache during the prefill phase is preserved to maintain the essential information, while a novel strategy based on sliding is proposed to select essential heavy hitters for the decoding phase. Memory usage and memory transfer are further optimized using adaptive and discontinuous strategies. Extensive experiments on LongGenBench show the effectiveness and generalization of SCOPE and its compatibility as a plug-in to other prefill-only KV compression methods.

An Automated Explainable Educational Assessment System Built on LLMs

In this demo, we present AERA Chat, an automated and explainable educational assessment system designed for interactive and visual evaluations of student responses. This system leverages large language models (LLMs) to generate automated marking and rationale explanations, addressing the challenge of limited explainability in automated educational assessment and the high costs associated with annotation. Our system allows users to input questions and student answers, providing educators and researchers with insights into assessment accuracy and the quality of LLM-assessed rationales. Additionally, it offers advanced visualization and robust evaluation tools, enhancing the usability for educational assessment and facilitating efficient rationale verification. Our demo video can be found at https://youtu.be/qUSjz-sxlBc.

Direct Preference Optimization Using Sparse Feature-Level Constraints

The alignment of large language models (LLMs) with human preferences remains a key challenge. While post-training techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have achieved notable success, they often introduce computational inefficiencies and training instability. In this paper, we propose Feature-level constrained Preference Optimization (FPO), a novel method designed to simplify the alignment process while ensuring stability. FPO leverages pre-trained Sparse Autoencoders (SAEs) and introduces feature-level constraints, allowing for efficient, sparsity-enforced alignment. Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence by using the feature-level offline reference. Experimental results on benchmark datasets demonstrate that FPO achieves a 5.08% absolute improvement in win rate with much lower computational cost compared to state-of-the-art baselines, making it a promising solution for efficient and controllable LLM alignments.

AERA Chat: An Interactive Platform for Automated Explainable Student Answer Assessment

Generating rationales that justify scoring decisions has emerged as a promising approach to enhance explainability in the development of automated scoring systems. However, the scarcity of publicly available rationale data and the high cost of annotation have resulted in existing methods typically relying on noisy rationales generated by large language models (LLMs). To address these challenges, we have developed AERA Chat, an interactive platform, to provide visually explained assessment of student answers and streamline the verification of rationales. Users can input questions and student answers to obtain automated, explainable assessment results from LLMs. The platform's innovative visualization features and robust evaluation tools make it useful for educators to assist their marking process, and for researchers to evaluate assessment performance and quality of rationales generated by different LLMs, or as a tool for efficient annotation. We evaluated three rationale generation approaches on our platform to demonstrate its capability.

GARLIC: LLM-Guided Dynamic Progress Control with Hierarchical Weighted Graph for Long Document QA

In the past, Retrieval-Augmented Generation (RAG) methods split text into chunks to enable language models to handle long documents. Recent tree-based RAG methods are able to retrieve detailed information while preserving global context. However, with the advent of more powerful LLMs, such as Llama 3.1, which offer better comprehension and support for longer inputs, we found that even recent tree-based RAG methods perform worse than directly feeding the entire document into Llama 3.1, although RAG methods still hold an advantage in reducing computational costs. In this paper, we propose a new retrieval method, called LLM-Guided Dynamic Progress Control with Hierarchical Weighted Graph (GARLIC), which outperforms previous state-of-the-art baselines, including Llama 3.1, while retaining the computational efficiency of RAG methods. Our method introduces several improvements: (1) Rather than using a tree structure, we construct a Hierarchical Weighted Directed Acyclic Graph with many-to-many summarization, where the graph edges are derived from attention mechanisms, and each node focuses on a single event or very few events. (2) We introduce a novel retrieval method that leverages the attention weights of LLMs rather than dense embedding similarity. Our method allows for searching the graph along multiple paths and can terminate at any depth. (3) We use the LLM to control the retrieval process, enabling it to dynamically adjust the amount and depth of information retrieved for different queries. Experimental results show that our method outperforms previous state-of-the-art baselines, including Llama 3.1, on two single-document and two multi-document QA datasets, while maintaining similar computational complexity to traditional RAG methods.