Do LLMs Know What to Ask and When? Evaluating Multi-Turn Information Seeking

When a user question is underspecified, a capable model should recognize that its context is insufficient, identify the missing information, ask for it, and respond only once that information determines a unique answer. We formalize multi-turn information seeking as solving a k-underspecified constraint satisfaction problem, where k is the number of variables jointly required to determine the target and therefore measures the degree of missing information. We instantiate the formulation in MT-InfoSeek, a controlled evaluation suite of 5,251 problems and 9,006 task instances spanning mathematics, logic, biology, medicine, and general knowledge. We evaluate models along three axes: what they ask, when they ask it, and how the acquired information affects the final answer. Performance degrades across models and domains as underspecification increases. Models recognize that additional information is needed but underestimate how much, and in logical problems at k = 2 they under-predict the degree of missing information about four times as often as they over-predict it. They also fail to identify a minimal sufficient set of queries, improve only marginally when given the true k, and often stop before acquiring sufficient information. In tasks with ordered dependencies, an incorrect query order reduces final accuracy even when the model eventually acquires all necessary information. We measure information seeking directly through final sufficiency, which records whether the acquired information determines the target independent of answer generation. This separation shows differences between models that final accuracy alone does not capture, and indicates that the ability to seek information over multiple turns is distinct from the ability to generate answers and is not measured by current LLM evaluations.

Learning to Grow Artificial Hippocampi in Vision Transformers for Resilient Lifelong Learning

Lifelong learning without catastrophic forgetting (i.e., resiliency) possessed by human intelligence is entangled with sophisticated memory mechanisms in the brain, especially the long-term memory (LM) maintained by Hippocampi. To a certain extent, Transformers have emerged as the counterpart ``Brain" of Artificial Intelligence (AI), and yet leave the LM component under-explored for lifelong learning settings. This paper presents a method of learning to grow Artificial Hippocampi (ArtiHippo) in Vision Transformers (ViTs) for resilient lifelong learning. With a comprehensive ablation study, the final linear projection layer in the multi-head self-attention (MHSA) block is selected in realizing and growing ArtiHippo. ArtiHippo is represented by a mixture of experts (MoEs). Each expert component is an on-site variant of the linear projection layer, maintained via neural architecture search (NAS) with the search space defined by four basic growing operations -- skip, reuse, adapt, and new in lifelong learning. The LM of a task consists of two parts: the dedicated expert components (as model parameters) at different layers of a ViT learned via NAS, and the mean class-tokens (as stored latent vectors for measuring task similarity) associated with the expert components. For a new task, a hierarchical task-similarity-oriented exploration-exploitation sampling based NAS is proposed to learn the expert components. The task similarity is measured based on the normalized cosine similarity between the mean class-token of the new task and those of old tasks. The proposed method is complementary to prompt-based lifelong learningwith ViTs. In experiments, the proposed method is tested on the challenging Visual Domain Decathlon (VDD) benchmark and the recently proposed 5-Dataset benchmark. It obtains consistently better performance than the prior art with sensible ArtiHippo learned continually.