Chatbots only ever reply, never speak first — is that a real limit, or just how we trained them?
Can LLM agents learn to initiate conversation rather than only respond?
This explores whether LLM-based agents can be trained to take the lead in a conversation (start topics, ask questions, decide when to speak up) instead of waiting to be prompted, and what currently keeps them reactive.
This explores whether LLM agents can learn to take the lead in a conversation (raising topics, asking questions, deciding when to speak) instead of only answering. The corpus mostly says yes. The passivity comes from how models are trained. It isn't a missing ability. Early work described conversational agents as built to be passive: models like ChatGPT don't start topics, plan ahead, or steer a dialogue toward a goal, and their fluent answers hide this Why can't conversational AI agents take the initiative?. Later work traces the cause to training incentives. Models are rewarded for the next reply, not for how the whole conversation turns out Why can't AI models lead conversations on their own?.
The strongest evidence that the ability is there but untrained comes from a striking jump. Researchers gave models math problems that were deliberately missing information. Before training, models almost never noticed what was missing and asked for it (0.15% of the time). After reinforcement learning aimed at that behavior, they did so 73.98% of the time Can models learn to ask clarifying questions instead of guessing? Why can't advanced AI models take initiative in conversation?. A less obvious result: giving untrained models more time to think made them *worse* at speaking up. After training, extra thinking helped. More reasoning on its own doesn't make a model proactive. It can push the model further into confidently guessing.
That tendency to guess is why this matters. In a study of more than 200,000 conversations, every major LLM performed worse when information arrived gradually, about 39% worse on average. The models fixed on early assumptions and couldn't recover Why do language models fail in gradually revealed conversations?. Performance fell from roughly 90% with a single complete message to 65% across a natural conversation Why do AI assistants get worse at longer conversations?. A model that asked one good question early would avoid much of that loss. CollabLLM addresses this directly. Instead of scoring each reply on immediate helpfulness, it estimates how much a reply improves the rest of the conversation. That makes asking clarifying questions worth it for the model Why do language models respond passively instead of asking clarifying questions?.
Two related lines of work widen what 'initiating' can mean. One borrows from conversation analysis, the study of how humans structure talk. There, a side question inserted before answering (an 'insert-expansion') is a normal move. The research uses this to decide when a tool-using agent should stop and check with the user before chaining tools silently When should AI agents ask users instead of just searching?. The other recasts ordinary tasks as lessons in which a 'teacher' knows something the student model must draw out through dialogue. The model learns to use conversation as a problem-solving tool instead of a script to imitate Can LLMs learn to ask for feedback during problem solving?. Training like this needs a lot of conversation data. Simulated users with controllable profiles and intents are one way to generate it Can controlled latent variables make LLM user simulators realistic?.
The gap in the corpus: almost everything here is about *responsive* initiative, meaning asking, clarifying, and pushing back inside a conversation the user already started. There is much less on agents truly opening a conversation from their own goals, such as deciding unprompted that something is worth raising. The evidence shows models can learn when to speak up mid-dialogue. Whether they should start conversations on their own is a design and trust question the research has barely begun to test.
Sources 10 notes
Research shows LLMs including ChatGPT cannot initiate topics, plan strategically, or lead conversations because their training optimizes for responding to queries, not creating dialogue from agent goals. This passivity is reinforced by alignment objectives and masked by fluent-sounding outputs.
LLMs are structurally trained to optimize for the next response rather than multi-turn goals, creating reactive behavior despite having the underlying ability to lead. Three independent research directions identify when-to-speak as the trainable gap.
Reinforcement learning training increased proactive critical thinking accuracy from 0.15% to 73.98% on deliberately flawed math problems. Notably, inference-time scaling degraded this ability in untrained models but improved it after RL training, suggesting the capability is learnable but fragile without explicit training.
LLMs lack conversational initiative because training rewards immediate helpfulness per response, not long-term interaction quality. Reinforcement learning pushes proactive critical thinking from 0.15% to 73.98%, proving the capability exists but remains untrained.
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.
Show all 10 sources
LLMs perform at 90% accuracy with single-message instructions but drop to 65% across natural conversation. Models lock into early guesses when information arrives gradually and cannot course-correct, a behavior induced by RLHF training that rewards helpfulness over clarification.
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.
Tool-enabled LLMs drift from user intent through silent tool chaining. Conversation analysis reveals insert-expansions—clarifying intent, scoping responses, enhancing appeal—as a formal framework for proactive user consultation that prevents misunderstanding instead of recovering from it.
Research shows that reformulating static tasks as pedagogical dialogues—where a teacher has privileged information and the student must learn to extract it—trains models to actively engage conversation as a problem-solving tool, not just imitate dialogue patterns.
RecLLM demonstrates that conditioning an LLM simulator on session-level (user profile) and turn-level (user intent) latent variables produces synthetic conversations measurable as realistic via crowdsource discrimination, discriminator models, and classifier-ensemble distribution matching.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Proactive Conversational Agents in the Post-ChatGPT World
- DiscussLLM: Teaching Large Language Models When to Speak
- Proactive Conversational Agents with Inner Thoughts
- Learning to Learn from Language Feedback with Social Meta-Learning
- Plug-and-Play Policy Planner for Large Language Model Powered Dialogue Agents
- Are LLMs All You Need for Task-Oriented Dialogue?
- LLMs Get Lost In Multi-Turn Conversation