Evgeny Kharlamov

A staggering 42.5% of correct answers in long-document VQA can't be derived from terminal working memory alone, revealing a crucial oversight in current evaluation methods.

Knowledge graph memory can achieve superior factual recall without the token costs of traditional retrieval methods, but its effectiveness hinges on overcoming the challenge of adapter selection.

A unified taxonomy reveals how diverse memory mechanisms in LLMs can be systematically understood and leveraged for future innovations.

MADA-RL boosts compact model reasoning accuracy by 2% with 16 times fewer trainable parameters, redefining how critics learn from generators.