Articulation Upper Bounds for Language Models

Alexander Spangher, Daniel E. Ho, Diyi Yang, Noah Goodman, Sanmi KoyejoarXiv 2026 · arXiv

TL;DR

Researchers at Stanford have developed a framework to measure the fundamental limits of what Large Language Models (LLMs) can communicate through language. They discovered that while most concepts become easier to transmit as models scale, a specific class of "socially anchored" concepts (like "comic voice") hits a hard ceiling that scaling alone doesn't fix—even though humans can often agree on them.

What Problem Does This Paper Address?

Language is a lossy channel. As Ludwig Wittgenstein famously asked, "Describe the aroma of coffee—why can’t it be done? Do we lack the words?" For LLMs, which perceive and act entirely through language, this isn't just a philosophical puzzle; it’s a technical bottleneck. If a concept cannot be clearly articulated in a prompt or rubric, we cannot reliably align an AI agent to it.

The authors set out to find the "Articulation Upper Bounds" of LLMs: the point where adding more words or more parameters no longer improves how well a concept can be transmitted from one model (the sender) to another (the receiver).

Methods

To find these limits, the team introduced two primary novel tools:

  1. $\epsilon$-estimation (Prompt Optimization): To ensure they were testing the best possible version of a message, they developed an algorithm that breaks prompts into "semantic units" (minimal fragments that change behavior). Using a "capture-recapture" technique borrowed from ecology—essentially counting how many unique "species" of prompt fragments appear across multiple runs—they can statistically estimate the performance of a theoretical "optimal" prompt that includes every possible useful instruction.
  2. The Reconstruction Bottleneck: This is a "language game" used to measure transmission. A "Transmitter" LLM uses a prompt to label a set of texts. A "Receiver" LLM sees only the texts and the labels and must try to "re-derive" the original prompt. If the Receiver’s recovered prompt can accurately label new, unseen data in the same way the Transmitter did, the concept has been successfully articulated.
The Reconstruction Bottleneck process: (1) Sender labels text, (2) Receiver reconstructs the prompt, (3) Mutual Information (MI) is calculated to measure success.

Datasets

The researchers evaluated 1,270 concepts across eight qualitative domains, including humor, creative writing, peer review, and code review. They also validated their $\epsilon$-estimation optimizer on six standard benchmarks (like HotpotQA and AIME) to prove it could outperform state-of-the-art prompt optimizers.

Key Insights and Ideas

  • The Three Regimes of Articulation: Concepts fall into three distinct categories based on how they scale with model size:
  • Reaches: Concepts that are already fully articulable (often local patterns like "joke ends with a surprise").
  • Rising: Concepts that aren't perfect yet but show clear improvement as models get bigger (often structural patterns like "worldbuilding coherence").
  • Bounded: Concepts that hit a plateau well below the "ground truth" ceiling.
Scaling curves for the three regimes: Rising (still climbing), Reaches (attained ceiling), and Bounded (plateaued below the limit).
  • Bounded Concepts are Socially Anchored: The concepts LLMs struggle most to articulate are those tied to identity and social-group awareness (e.g., "persona," "in-jokes," or "community stance"). These require "collective tacit knowledge" that current models seem unable to pass through the bottleneck of a text prompt.
  • The "Executor Relative" Limit: Interestingly, humans have high levels of agreement on 76% of the concepts that LLMs find "inarticulable." This suggests the limit isn't in the language itself, but in the "executors" (the LLMs). Our current models lack the shared social context humans use to fill in the gaps of language.
  • Showing vs. Telling: The researchers hypothesized that "tacit" knowledge might be transmitted via examples (in-context learning) even if definitions fail. Surprisingly, they found that for bounded concepts, providing examples rarely helped more than a simple definition.

Key Results

  • Optimizer Performance: The $\epsilon$-estimation algorithm outperformed existing prompt optimizers (like GEPA) by up to 14% on standard benchmarks.
  • Scaling Success: Most concepts (roughly 70%) are in the "Rising" regime, suggesting that future, more powerful models will continue to get better at understanding complex structural requirements.
  • The Social Ceiling: Only about 5% of concepts (65 out of 1,270) were found to be truly "Bounded." However, these are often the very concepts we care about for high-level alignment, such as "comic voice" or "professional tone."
The distribution of concept types across the three regimes. Bounded concepts are heavily skewed toward identity and social-group awareness.

Limitations and Future Work

The authors acknowledge that their "reconstruction" measure is unsupervised; it rewards consistency between models, which isn't always the same as "truth." Furthermore, the study focused on creative and qualitative domains; it’s unclear if factual or procedural concepts (like complex legal reasoning) face similar social-group boundaries.

The findings suggest a major fork in the road for AI research: if we want models to understand these "bounded" concepts, we may need to move beyond just scaling up text-based training and explore novel learning mechanisms that can capture the social context humans use to understand the "unsaid."

Summary updated

Abstract

Language captures only a subset of the perceptions, intuitions and intents that form meaning between speakers. What, then, are the limits of language? We take steps towards measuring this in LLMs as follows: for a set of concepts, we measure how well each concept can be transmitted between two LLMs as their capabilities scale. This requires novel measurement tools: (1) optimal messages: we introduce a prompt optimization algorithm that beats existing algorithms by up to 14% on benchmarks and accurately estimates prompt upper bounds. (2) A transmission measure: the sender is asked to label text; the receiver then tries to recover the sender's prompt, and recovery is measured; we find this correlates strongly with human label error. With these measures, we use observational scaling laws to study how transmissibility of a concept, conveyed with optimal messages, increases as transmitter and receiver capacity grows: does it asymptote below or above a threshold? Measuring 1.2k concepts in eight domains, we find concepts that (a) can be fully articulated; these concepts usually depend on local patterns, e.g. "joke ends with surprise"; (b) might be articulated by future, stronger models; these rely on whole-text structural patterns, e.g. "story's worldbuilding coherence"; (c) concepts that face articulability limits; these concepts' meaning, we find, is tied to identity and social-group awareness, e.g. "comic voice". Interesting, these same concepts are not all challenging for humans: humans have moderate-to-high levels of agreement in 76% of concepts that LLMs struggle with; articulation limits, we find, are executor relative. Finally, we characterize the nature of improvement in each case: while recoverable concepts can be articulated through definitions and explanations, irrecoverable concepts cannot be transmitted, even with examples, and might require novel learning mechanisms.
BibTeX
@misc{spangher2026articulation,
  author = {Alexander Spangher and Daniel E. Ho and Diyi Yang and Noah Goodman and Sanmi Koyejo},
  title = {{Articulation Upper Bounds for Language Models}},
  year = {2026},
  eprint = {2610.19721},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.19721},
}
Figure