Why Learning from Context Is Harder Than We Thought

The MLNLP academic Talk on March 7, 2026 featured Dou Shihan, a PhD candidate at Fudan University whose research focuses on post‑training of large language models.

Recent years have seen rapid advances in language models on knowledge reasoning, mathematics, and programming, but these benchmarks mainly assess the model’s ability to retrieve and apply stored knowledge. When a task demands that the model learn new facts from the current input context and reason on that basis, performance drops sharply.

In real‑world scenarios, both applications and humans rely heavily on quickly extracting information from documents, rules, logs, or environments and using it immediately. Existing models, however, tend to depend on parametric knowledge; even when relevant context is provided, they may ignore or misuse it, indicating that they have not truly learned “how to learn from context.”

The analysis identifies several systematic biases that hinder context learning in language models and proposes a possible improvement direction: synthesizing rubrics to guide the model toward better understanding and utilization of contextual information.