Clinical diagnosis is inherently sequential: clinicians escalate from cheap to costly tests only when additional evidence is expected to resolve diagnostic uncertainty. We present ActiveMedAgent, a framework that brings this cost-aware sequential logic to multimodal medical AI. Given a frozen, API-accessed vision-language model, ActiveMedAgent probability distributions over candidate diagnoses induced each acquisition by its per-step diagnostic utility minus cost. A lightweight MLP controller is then trained offline on these scored trajectories, learning when to request additional evidence and when to commit. Across three commonly used benchmarks, trajectory-based policy learning consistently outperforms both unguided acquisition and full-modality baselines. Notably, we identify an information overload effect. In 175 cases, the agent produces a correct diagnosis with fewer channels while the full-modality baseline fails, showing that learning what to omit can be as important as learning what to acquire

Multimodal large language models (MLLMs) have shown strong performance in perception and question-answering, but many clinical decisions require reasoning across observations collected over time rather than from a single input. We study this problem in longitudinal modeling, where evidence evolves across visits. We introduce LUMINA (LongitUdinal MultImodal reasoNing Agent), a single-agent framework that incrementally builds an auditable evidence state before making predictions. Instead of predicting directly from all observations, LUMINA organizes current evidence, compares observations across prior visits, maintains longitudinal memory, optionally retrieves relevant cross-subject summaries, and verifies whether the accumulated evidence supports the final decision. We evaluate LUMINA on two subject-level benchmarks: Alzheimer’s disease classification and 36-month progression forecasting from observed visit prefixes. Across nine MLLM backbones, LUMINA consistently outperforms MLLM baselines and an adapted medical reasoning agent, achieving 13.8%–18.5% relative balanced accuracy gains on classification and 16.8%–24.8% gains on progression forecasting over the strongest non-agent baseline. Additional evaluations preserve the same method ordering on Parkinsonian-state classification and breast cancer treatment-response prediction.

Lifelong editing of LLMs requires storing thousands of edits after acquisition. A widely used family of approaches attaches one LoRA adapter per edit, which preserves behavior but grows linearly in storage. To address this challenge, we propose LadderEdit, a method that compresses each LoRA adapter after it is acquired. Each edit is first stored at low rank as a cheap sketch. We then check whether this sketch still satisfies the rewrite, generalization, and locality contract on probe prompts. Edits that pass keep the sketch; those that fail are promoted to a higher rank along a ladder until the contract is met. Because every edit retains some representation, coverage is maintained, and only hard edits consume more rank. Across ZsRE, CounterFact, and WikiBigEdit benchmarks on LLaMA-3-8B, Mistral-7B, and Qwen2.5-7B, LadderEdit tracks exact LoRA storage at 5.2× less memory and remains effective at 50,000 sequential edits.