Taeyang Yoon

Ji-jun Park, Soo-joon Choi, Jiwon Jeong et al.

Large language models (LLMs) frequently encode factual and reasoning knowledge in their internal representations that is not faithfully reflected in their surface-level outputs -- a phenomenon known as \emph{latent knowledge}. Existing approaches to eliciting latent knowledge, such as Contrastive Consistency Search (CCS), rely on contrastive activation patterns and struggle with complex multi-step reasoning tasks, while mechanistic interpretability tools have primarily been used to \emph{understand} model behavior rather than to \emph{extract} hidden knowledge. We present \textbf{MechELK}, a unified three-stage framework that bridges mechanistic interpretability and latent knowledge elicitation. MechELK operates through: (1) \textbf{Locate} -- using Sparse Autoencoder (SAE) feature analysis and activation patching to identify knowledge-bearing representations; (2) \textbf{Verify} -- employing causal probing to distinguish genuine latent knowledge from spurious correlations; and (3) \textbf{Elicit} -- applying representation engineering to surface hidden knowledge without modifying model weights. Evaluated on TruthfulQA, a curated Deceptive Alignment benchmark, and the Quirky LM dataset, MechELK achieves an average elicitation accuracy of 84.7\%, outperforming CCS by 6.2\% and direct linear probing by 9.1\%. Crucially, MechELK successfully identifies latent knowledge in 78.3\% of cases where the model's surface output is incorrect or evasive, demonstrating its utility for AI safety applications including deceptive alignment detection.

Seungmin Han, Haeun Kwon, Ji-jun Park et al.

Despite significant advancements in Large Language Models (LLMs) and Large Vision-Language Models (LVLMs), current models still face substantial challenges in handling complex, multi-turn, and visually-grounded tasks that demand deep reasoning, sustained contextual understanding, entity tracking, and multi-step instruction following. Existing benchmarks often fall short in capturing the dynamism and intricacies of real-world multi-modal interactions, leading to issues such as context loss and visual hallucinations. To address these limitations, we introduce MMDR-Bench (Multi-Modal Dialogue Reasoning Benchmark), a novel dataset comprising 300 meticulously designed complex multi-turn dialogue scenarios, each averaging 5-7 turns and evaluated across six core dimensions including visual entity tracking and reasoning depth. Furthermore, we propose CoLVLM Agent (Contextual LVLM Agent), a holistic framework that enhances existing LVLMs with advanced reasoning and instruction following capabilities through an iterative "memory-perception-planning-execution" cycle, requiring no extensive re-training of the underlying models. Our extensive experiments on MMDR-Bench demonstrate that CoLVLM Agent consistently achieves superior performance, attaining an average human evaluation score of 4.03, notably surpassing state-of-the-art commercial models like GPT-4o (3.92) and Gemini 1.5 Pro (3.85). The framework exhibits significant advantages in reasoning depth, instruction adherence, and error suppression, and maintains robust performance over extended dialogue turns, validating the effectiveness of its modular design and iterative approach for complex multi-modal interactions.