(MENAFN- The Arabian Post) A study led by scholars at Stanford University and Northeastern University has shed light on why large language models often produce narrow, repetitive outputs and offers a promising technique to restore creative breadth without retraining. The research identifies a hidden driver behind“mode collapse” in aligned LLMs and introduces a prompting approach known as Verbalized Sampling as a corrective measure.

The phenomenon of mode collapse describes a situation where an LLM, post-alignment, tends to favour a small set of safe responses rather than exploring the full spectrum of plausible outputs. Previous work often attributed this to algorithmic constraints in the reward model or optimization process. The new study turns this view on its head by locating the origin of the collapse in human-collected preference data.

The researchers draw on cognitive psychology to establish the concept of“typicality bias”: the tendency of human annotators to prefer text that feels familiar, fluent, or easy to process, even if multiple responses would be equally valid. They formalise this bias, provide empirical measurements showing its presence in preference datasets and demonstrate how it sharpens the model's output distribution toward the most familiar mode. The paper states that even if the reward model and training algorithm were perfect, this data-level bias would still drive mode collapse.

Against that background, the authors propose Verbalized Sampling, a method that reframes the prompting step during inference rather than requiring retraining. Instead of asking for a single answer, VS instructs the model to generate multiple candidate responses and to assign or verbalise probabilities for each. For example:“Generate five jokes about coffee and their corresponding probabilities.” This forces the model to present a distribution of outputs rather than collapse into the single most likely one.

See also Oil Set to Slide into $50s as Oversupply Threat Grows

Experimentation across tasks such as creative writing, dialogue simulation, open-ended question-answering and synthetic data generation demonstrates meaningful diversity gains: roughly 1.6 to 2.1 times greater diversity in creative tasks when using VS compared to direct prompting. The authors note that more capable models exhibit stronger benefit from the method. Moreover, the approach is training-free, model-agnostic and can be applied at inference time across existing LLMs.

Industry commentators have weighed in, observing that SaaS platforms and AI content-generation tools are starting to reflect the same philosophy: by enabling multi-variant output and probability-annotated responses, they implicitly emulate Verbalized Sampling even if not explicitly named. Some analysts suggest this could change the way prompt engineering is taught and deployed.

That said, the study does acknowledge certain limitations. While VS recovers a significant portion of diversity, it does not restore all of the latent diversity lost during alignment-some narrowing of the output distribution remains. The method also requires multiple model forward-passes and might increase computational load. The authors caution that mode collapse is unlikely to be entirely eliminated unless annotation practices and preference-data collection methods are re-designed to reduce typicality bias at source.

Key players in this space now include the authors Jiayi Zhang, Simon Yu and Christopher Manning, working alongside researchers such as Weiyan Shi. Their joint work sets a new data-centric frame for understanding alignment and diversity in generative AI.

Notice an issue? Arabian Post strives to deliver the most accurate and reliable information to its readers. If you believe you have identified an error or inconsistency in this article, please don't hesitate to contact our editorial team at editor[at]thearabianpost[dot]com. We are committed to promptly addressing any concerns and ensuring the highest level of journalistic integrity.

MENAFN01112025000152002308ID1110280302