Can an AI learn to ramble when a problem is hard and keep it short when it's not?
Can AI agents be trained to adjust response length based on context?
This explores whether AI agents can learn to make their output longer or shorter depending on the situation, rather than producing a fixed amount of text no matter what is asked. Here 'length' covers both how much the model thinks before answering and how much it says.
This explores whether AI agents can learn to make their output longer or shorter depending on the situation. The corpus says yes, though it rarely uses the phrase 'response length.' Most of the relevant work frames it as a different question: when should a model think hard, when should it answer fast, and when should it say less and ask more?
The most direct evidence is about reasoning length. Thinkless trains one model to choose between a long chain of reasoning and a short, direct answer Can models learn when to think versus respond quickly?. Nobody tells the model which questions are hard. It works this out through reinforcement learning. The useful trick is that the training treats the choice of mode ('think or not?') separately from improving the answer itself. Without that split, training tends to collapse into always using one mode. In other words, the hard part is less teaching a model to be brief than stopping training from flattening its judgment about when brevity fits.
Length also isn't only about word count. In conversation, the right length depends on what the other person needs, and sometimes the best reply is short: a clarifying question. Standard RLHF rewards the reply that looks most helpful right now, which pushes models toward long, confident answers built on guesses Why do language models respond passively instead of asking clarifying questions?. That habit is one reason assistants get worse over longer chats. They commit early to a long answer and can't recover Why do AI assistants get worse at longer conversations?. Training can fix this. RL raised one model's rate of spotting missing information from almost zero to about 74% Can models learn to ask clarifying questions instead of guessing?. In emotional-support chat, agents that adjusted their pacing to the moment were rated as better listeners, and users shared more about themselves Can pacing strategies make conversational agents feel more human?. Timing can be learned too. A single streaming model can pick up when to speak and when to stay quiet as part of ordinary training Can a single model learn when to speak and respond?.
You don't always need to retrain at all. Microsoft's ACON improves written compression guidelines by studying cases where agents failed. It cut agents' peak token use by 26–54% without changing any model weights Can prompt optimization cut agent token usage without retraining?. Practitioners add a related point: rules about tone and behavior, including how much to say, belong in the agent's instructions, not mixed in with factual content Should factual content and behavior rules live separately?.
The takeaway you might not expect: the default bias toward long answers is mostly created by how models are rewarded during training. Once the reward pays off over a whole conversation rather than one turn, a model can learn that the right length is sometimes a single question. The corpus has no study that directly trains for matching output word count to context, so the evidence here comes from these neighboring problems.
Sources 8 notes
Thinkless trains a single model to select between extended reasoning and direct responses using DeGRPO, which decouples mode selection from answer refinement. This prevents mode collapse and enables self-calibrated routing without explicit difficulty labels.
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.
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.
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.
A 50-person study found that conversational agents using dynamic pacing strategies outperformed static-pacing baselines on perceived listening quality, human-likeness, and user self-disclosure in text-based career and relationship support scenarios.
Show all 8 sources
Wan-Streamer represents language, audio, and video as one interleaved causal token stream, allowing response timing and turn management to be learned jointly within a single Transformer rather than engineered as separate modules, achieving sub-second latency.
Microsoft's ACON method reduces agent token consumption and improves task success by iteratively refining compression guidelines through failure analysis, not weight updates. This approach works across proprietary LLM agents and costs under $2 per benchmark.
Fin's Blueprint Team recommends routing factual content to retrievable knowledge bases while keeping only tone and behavioral directives in agent instructions. This split aligns with instruction-tuning research showing models follow what-to-do rules better than what-to-ignore constraints.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Can Large Language Models Reason and Optimize Under Constraints?
- Proactive Conversational Agents with Inner Thoughts
- Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation
- Proactive Conversational Agents in the Post-ChatGPT World
- Knowing but Not Showing: LLMs Recognize Ambiguity but Rarely Ask Clarifying Questions
- Learning to Learn from Language Feedback with Social Meta-Learning
- DiscussLLM: Teaching Large Language Models When to Speak
- LLMs Get Lost In Multi-Turn Conversation