Can language models produce language more efficiently through interaction?
This reads the question as: does back-and-forth interaction actually make a language model better at producing language — or does interaction mostly expose limits that single-shot prompting hides?
This explores whether interaction is a performance multiplier for language models or a stress test that reveals their seams — and the corpus leans surprisingly hard toward the second reading, with one important exception. The default story most people assume is that more turns mean more context, and more context means better output. The collection mostly says the opposite happens by accident. When a task is revealed gradually across turns, all major models show a roughly 39% performance drop because they lock onto a premature guess early and can't recover, with agent-style fixes clawing back only 15–20% of the loss Why do language models fail in gradually revealed conversations?. So interaction can make a model *worse*, not because it runs out of information but because it commits too soon.
A deeper reason is that today's models aren't trained to treat interaction as collaborative work. Standard RLHF optimizes for the reward on the *next* turn, which quietly teaches a model to be immediately agreeable rather than to ask the clarifying question that would make the whole exchange land better Why do language models respond passively instead of asking clarifying questions?. The encouraging counterpoint sits right next to it: when you change the reward to estimate the long-term value of an interaction instead of the next reply, models start actively discovering intent. That's the strongest 'yes' in the corpus — interaction *can* produce better language, but only once the training objective is built for multiple turns rather than for one.
The harder limits are structural, not just about reward design. Human conversation runs on implicit maintenance — repairing references, handing off topics, keeping things smooth — and models don't develop these because training rewards predicting information, not the relational glue that holds a conversation together Why don't language models develop conversation maintenance skills?. Worse, a model can't symmetrically *update* shared assumptions the way two people do: it interprets every later turn inside the frame of the opening prompt, so when you pivot or contradict yourself, you end up being the only one keeping the shared scoreboard Can LLMs truly update shared conversational common ground?. Interaction in the human sense requires both sides to revise common ground; the model mostly can't.
There's also a ceiling argument worth knowing about: interaction with *itself* won't rescue a model. Self-improvement is formally bounded by a generation-verification gap, meaning every reliable fix needs something external to validate it — a model can't talk its way past its own limits through more internal back-and-forth What stops large language models from improving themselves?. That reframes 'efficiency through interaction' usefully: the gains come from interaction with a *grounded outside party* (a user, a verifier, a tool), not from more turns in isolation.
So the answer the collection leaves you with is sharper than the question: interaction doesn't automatically make language production more efficient, and naively it often degrades it. The lever that actually works is reframing interaction as a long-horizon objective with an external signal — train for the value of the whole exchange, give the model something outside itself to check against, and the multi-turn drop can start to invert into a gain.
Sources 5 notes
Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.
CollabLLM demonstrates that standard RLHF training optimizes for immediate helpfulness, discouraging models from asking clarifying questions or offering multi-turn insights. Multi-turn-aware rewards that estimate long-term interaction value enable active intent discovery and genuine collaboration.
Humans keep conversations smooth through implicit techniques like reference repair and topic hand-off that sustain relational interaction, not convey information. Language models don't develop these because training signals reward information prediction, not relational work.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
Self-improvement in LLMs is formally bounded by the generation-verification gap, meaning every reliable fix requires something external to validate and enforce it. Models cannot escape this constraint through metacognition alone.