LSD for MLLMs: Reinforcement Learning Cuts the Demo Fat

Multimodal Large Language Models (MLLMs) have spent the last two years drowning in their own demo debt. The standard fix—k-Nearest Neighbor (kNN) search—prioritizes similarity over substance, churning out redundant examples that flatten the output range of complex tasks like factual regression. Enter Learning to Select Demonstrations (LSD), a reinforcement learning approach that reframes demo selection as a sequential decision problem.

Instead of letting kNN lazily grab the nearest neighbors, LSD trains a Dueling Deep Q-Network (DQN) with a query-centric Transformer Decoder to construct optimal sets. The goal isn’t just to pick similar examples—it’s to pick the ones that actually teach the model something new.

The paper’s abstract, posted in March 2026, reads like a direct critique of the status quo. kNN’s redundancy isn’t just inefficient; it’s actively harmful for tasks where output diversity matters. LSD’s RL-based policy aims to maximize downstream performance, but the abstract stops short of sharing any numbers. That’s the first red flag—or at least the first question mark. For all the talk of ‘optimal’ demo sets, we’re still in the realm of theoretical improvement, not benchmarked gains. The original kNN approach it’s replacing was never designed for multimodal complexity, so the bar for ‘better’ isn’t exactly high.

The technical community has already started poking at the gaps. On GitHub discussions, developers note that RL-based selection isn’t new—it’s been tried in text-based ICL for years—but the multimodal twist is what’s drawing attention. The real test will be whether LSD can scale beyond visual tasks. The paper’s title hints at ‘visual in-context demonstrations,’ but the method’s architecture doesn’t seem tied to images. If it works, it could become a drop-in replacement for kNN across modalities.

So who stands to gain?

The obvious winners are the teams already invested in MLLMs for complex regression tasks—think autonomous systems, medical imaging, or any domain where output range matters more than raw similarity. Companies like Google DeepMind and Meta have been vocal about the limitations of kNN, but neither has shipped a production-ready alternative. LSD’s RL approach could fill that gap, assuming the performance claims hold up under scrutiny.

The competitive pressure isn’t just on the model developers, though. The entire ‘in-context learning’ narrative has been built on the back of cheap, unsupervised demo selection. If LSD proves that smarter selection leads to better performance, it could force a reckoning: either invest in RL-based curation or admit that your model’s ‘learning’ is just memorization in disguise. The Hugging Face community has already started debating whether this is a ‘nice-to-have’ or a ‘must-have’ for future MLLM architectures.

There’s also the question of implementation cost. kNN is fast and cheap; RL is neither. The paper’s Dueling DQN with a Transformer Decoder isn’t exactly lightweight, and training a policy to select demos adds another layer of complexity to an already expensive pipeline. For now, the trade-off is theoretical. Until someone runs the numbers on real-world tasks—and shares them publicly—LSD remains an intriguing idea, not a proven upgrade.

The real signal here isn’t the method itself, but the shift in thinking. Demo selection isn’t just a preprocessing step anymore; it’s a first-class problem. That’s the kind of reframing that often precedes real progress—even if the first attempt is more hype than substance.