# Mode Collapse Mode collapse is what happens when a generative model stops covering the full range of things it could produce and keeps producing the same few. A "mode" is a peak in a probability distribution: a common kind of output. Real data has many modes (cats, dogs, cars; formal emails, jokes, haikus). A collapsed model only visits one or a handful of them. Ever noticed that every chatbot writes the same joke about atoms "making up everything"? Or that AI-written LinkedIn posts all sound alike? That's the LLM version of it. ## The term comes from GANs The term comes from Generative Adversarial Networks (Goodfellow et al., NeurIPS 2014, arXiv:1406.2661). A GAN has two networks: a generator that creates fake samples, and a discriminator that tries to spot them. The generator only needs to fool the discriminator, not to cover the whole data distribution. So it can find ONE output that works and repeat it forever. The original paper called this the "Helvetica scenario", and Ian Goodfellow's NIPS 2016 tutorial (arXiv:1701.00160) named it mode collapse. There are degrees: - **Full collapse**: the generator produces essentially one output - **Partial collapse, or mode dropping**: some modes of the real distribution are simply never generated (a face GAN that never produces glasses, for example) Fixes included unrolled GANs (Metz et al., 2016) and minibatch discrimination (Salimans et al., 2016), which let the discriminator notice when a whole batch looks the same. ## RLHF narrows the distribution in LLMs In language models, the term is used more loosely for the narrowing that happens during [[AI Post-Training|post-training]], especially [[Reinforcement Learning From Human Feedback (RLHF)|RLHF]]: - **Kirk et al.** (*Understanding the Effects of RLHF on LLM Generalisation and Diversity*, arXiv:2310.06452, ICLR 2024) found that RLHF "generalises better than SFT to new inputs" but "significantly reduces output diversity compared to SFT across a variety of measures." A real trade-off - **Kadavath et al.** (Anthropic, arXiv:2207.05221) described the mechanism: RL "tends to collapse language model predictions towards behaviors that receive the most reward." That same collapse wrecks calibration, so diversity loss and calibration loss are two views of one phenomenon (see [[AI Model Calibration]]) - **Verbalized Sampling** (Zhang et al., arXiv:2510.01171, October 2025) points at the data: human annotators systematically prefer familiar, typical text (a "typicality bias" well known in cognitive psychology). So preference data itself pushes toward the mode, even with a perfect reward model. Their training-free fix is to ask for several responses with their probabilities and sample from that list, which increased diversity 1.6 to 2.1x in creative writing That last result tells me something important: the base model's diversity is still in there. Post-training suppresses it; it doesn't erase it. ## Mode dropping vs mode collapse (TypeSafe's distinction) [[TypeSafe AI]]'s primer is careful with words here. It says preference optimization causes **mode dropping**: "the model learns to favor a particular style, such as instruction following, while reducing the probability of other possible outputs." It calls that "a milder version of mode collapse", the full GAN failure where a generator produces "the same kind of output repeatedly." Diogo Almeida puts it bluntly on Latent Space: "GANs mode drop. They drop the minority classes and just do the really common ones." And he goes further: we "have dropped every mode other than autoregressive chat-tuned LLMs," i.e. the whole industry collapsed onto one shape of model. ## Why it matters for decisions For a chatbot, mode dropping means blander prose. For a classifier, it's much worse: it means the model systematically under-predicts rare classes. The rare fraud case, the unusual support ticket, the minority label. Those are often exactly the cases you built the classifier to catch. That's a big part of why [[Jev]] is trained differently. A [[Decision Models (DMs)|decision model]] needs its probability distribution over options to reflect reality, rare options included. A distribution squashed onto the "typical" answer is useless for that, however fluent the model sounds. See [[Reinforcement Learning for Calibrated Decisions (RLCD)|RLCD]]. ## Limits of the concept - The LLM usage is fuzzy. People say "mode collapse" for anything from repetitive jokes to low sampling diversity at temperature 0. Temperature ([[AI Temperature]]) changes how you sample; mode collapse is about the distribution itself - Less diversity isn't always bad. For a customer support bot, consistency is a feature. The problem is when narrowing hides uncertainty or deletes legitimate answers - It's hard to measure. Diversity metrics (distinct n-grams, embedding spread) disagree with each other ## My take I think this is one of the most practical concepts to understand if you use AI for writing. The "AI voice" everyone recognizes is partly mode collapse: the model gravitating to the most typical phrasing. It's also why prompting for variety ("give me 5 very different options") works better than asking the same question 5 times. And for decision systems, the lesson is sharper: if a model was trained to please the average rater, don't expect it to respect the minority class. ## References - [Generative Adversarial Nets (Goodfellow et al., arXiv:1406.2661)](https://arxiv.org/abs/1406.2661) - [NIPS 2016 Tutorial: Generative Adversarial Networks (arXiv:1701.00160)](https://arxiv.org/abs/1701.00160) - [Understanding the Effects of RLHF on LLM Generalisation and Diversity (arXiv:2310.06452)](https://arxiv.org/abs/2310.06452) - [Language Models (Mostly) Know What They Know (arXiv:2207.05221)](https://arxiv.org/abs/2207.05221) - [Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity (arXiv:2510.01171)](https://arxiv.org/abs/2510.01171) - [TypeSafe AI primer](https://docs.typesafe.ai/introduction/machine-learning-primer) - [Jev: System One models for Prod, not God, with Diogo Almeida (Latent Space)](https://www.latent.space/p/jev) ## Related - [[Reinforcement Learning From Human Feedback (RLHF)]] - [[AI Post-Training]] - [[AI Model Calibration]] - [[AI Sycophancy]] - [[AI Temperature]] - [[Reinforcement Learning for Calibrated Decisions (RLCD)]] - [[Deep Learning]] - [[Yoshua Bengio]] - [[Typicality Bias]]