The development of adaptive agents and foundation models marks a significant shift toward AI systems that can continually learn, adapt, and evolve in response to new information, changing environments, and user preferences. Current AI models are typically trained on static data, with limited ability to adapt through context post-deployment. Our goal is to enable agents to continuously absorb new knowledge and compress it into reusable representations for more up-to-date responses. This capability is also valuable for third-party customization, personalization, and safety alignment. We are interested in both the foundational study of sequential learning dynamics in large language models and practical applications that demand adaptive agents, such as personalized assistance, multimodal learning, and news forecasting.
CoRR · 2026-07-02
AdaJEPA adapts a latent world model inside closed-loop MPC, using each observed transition as a self-supervised signal before the next replan.
CoRR · 2026-05-26
A simple post-hoc calibrator that maps an LLM's verbalized point forecast to a Beta distribution over event probability, trained on binary outcomes and human forecasts.
PhilPapers preprint · 2026-04-08
We argue self-consciousness requires a learned self — bounded integration of experience produces a perspective that, under continuous order-sensitive learning, becomes a temporally extended identity that current AI systems lack.
CoRR · 2026-03-10
We introduce a new benchmark to evaluate multiple embodied AI agents collaborate to answer human queries from their past experiences.
CoRR · 2025-12-02
Cross-family verification is found to be especially effective, and post-training reduces self-improvement but strengthens cross-family improvement.
ICLR 2026 · 2025-10-08
Action-conditioned Root mean squared Q-Functions (ARQ) is a novel backprop-free value estimation method that applies a goodness function and action conditioning for local reinforcement learning.
CoRR · 2025-08-21
StreamMem is a query-agnostic KV cache memory mechanism for streaming video understanding.
ICML 2026 · 2025-07-06
Context Tuning directly optimizes an LLM's memory representation for efficient adaptation without updating model weights.
ICML 2025 · 2024-11-13
Our new benchmark, Daily Oracle, automatically generates question-answer (QA) pairs from daily news, challenging LLMs to predict "future" events based on pre-training data.
COLM 2024 · 2024-03-22
CoLLEGe is a meta-learning framework capable of generating flexible embeddings for new concepts using a small number of example sentences or definitions.
NeurIPS 2024 · 2024-03-14
We discover a curious and remarkable property of LLMs fine-tuned sequentially in this setting: they exhibit anticipatory behavior, recovering from the forgetting on documents before encountering them again.
ICML 2024 · 2023-12-20
We explore the behavior of LLMs finetuned on noisy custom data containing unsafe content and propose a simple filtering algorithm for detecting harmful content based on the phenomenon of selective forgetting.
CoRR · 2023-12-07
LifelongMemory is a new framework for accessing long-form egocentric videographic memory through natural language question answering and retrieval.
