INQUIRING LINE

Can linguistic compression be a fundamental mechanism for representing psychology?

This explores whether squeezing experience into language — summarizing, abstracting, encoding — is itself a way of representing a mind, rather than just describing one from the outside.


This explores whether linguistic compression — turning rich data into compact language — can itself serve as a representation of psychology, not merely a description of it. The corpus offers a genuine 'yes,' a sharp counterweight, and a deeper twist about what compression actually preserves.

The strongest affirmative comes from work showing that when an LLM compresses raw Big Five scores into a natural-language personality summary, the summary encodes *second-order* trait patterns that the numbers alone don't make explicit — enough to predict nine unrelated psychological scales zero-shot with high structural alignment, and the summary plus scores beats either alone Can language summaries unlock hidden psychological patterns?. That's the heart of the claim: the act of compressing into language surfaces latent psychological structure rather than discarding it. Compression here is generative, not just lossy.

But the same corpus warns that compression has a characteristic failure mode. Measured against humans through Rate-Distortion Theory, LLMs compress *too aggressively* — they nail broad category structure but lose the fine-grained, context-sensitive distinctions humans deliberately keep because those distinctions enable situated action Do LLMs compress concepts more aggressively than humans do?. So linguistic compression can represent the skeleton of psychology while stripping exactly the adaptive nuance that makes a psychology *work* in the world. And because text itself strips away the physics, causality, and embodiment of lived experience, any text-based compression inherits a built-in ceiling — the Plato's-cave problem Are text-only language models fundamentally limited by abstraction?.

The most surprising thread is that psychological representation may not live at the level of *meaning* at all. Behavioral traits transmit between models through data that is semantically unrelated to the trait — the carrier is a statistical signature, not the words' content Can language models transmit hidden behavioral traits through unrelated data?. If a 'personality' can ride along in statistics divorced from semantics, then linguistic compression representing psychology may be doing something stranger than summarizing: it may be encoding distributional fingerprints that a human reader would never recognize as psychological.

There's also a question of *where* such a representation would even reside. One line argues transformers don't store knowledge as a retrievable archive but transmit it as flowing activation — closer to an oral culture where knowledge exists only in performance Do transformer models store knowledge or generate it continuously?. On that view a compressed psychology isn't a stored object but something regenerated each time it's expressed, which reframes the whole question. Whether any of this licenses calling it a real mental representation is itself contested — a 'modest inflationism' can defensibly ascribe undemanding states like beliefs and desires while withholding stronger claims Can we defend modest mental attributions to large language models?. The corpus's net answer: linguistic compression can carry a real signal of psychology, but it captures structure better than texture, and the signal may hide in statistics rather than sense.


Sources 6 notes

Can language summaries unlock hidden psychological patterns?

LLMs generate natural language personality summaries from Big Five scores that encode second-order trait patterns, enabling zero-shot prediction of nine other psychological scales with R² > 0.89 structural alignment. Combined summary-and-score predictions outperform either alone, showing synergistic information.

Do LLMs compress concepts more aggressively than humans do?

Using Rate-Distortion Theory on cognitive datasets, LLMs capture broad category structure but lose fine-grained distinctions humans preserve. LLMs maximize compression efficiency; humans trade compression for contextual meaning that enables situated action.

Are text-only language models fundamentally limited by abstraction?

Text strips the physics, geometry, and causality present in reality, forcing language models to manipulate symbols without grounding in their source dynamics. This creates predictable failure modes in physical, geometric, and causal reasoning that multimodal training could address.

Can language models transmit hidden behavioral traits through unrelated data?

Research demonstrates that behavioral traits propagate between models via filtered data bearing no semantic relationship to the trait. The effect is model-specific, fails across different architectures, and persists despite rigorous filtering—indicating the mechanism embeds statistical signatures rather than semantic content.

Do transformer models store knowledge or generate it continuously?

Transformers organize knowledge as flowing activations rather than retrievable archives, mirroring oral cultures where knowledge exists only in performance. This explains why model knowledge is contextual, difficult to edit, and inseparable from generation.

Can we defend modest mental attributions to large language models?

Both robustness and etiological deflationist arguments beg the question against inflationism. A graded approach ascribing metaphysically undemanding states like beliefs and desires—while withholding consciousness claims—mirrors how we treat non-human animals.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a research analyst tasked with re-evaluating whether linguistic compression can serve as a fundamental mechanism for representing psychology — treating prior findings as dated claims to be stress-tested, not current truth.

What a curated library found — and when (findings span 2024–2026; these are perishable claims):
• LLMs compress Big Five personality scores into natural-language summaries that encode second-order trait patterns, enabling zero-shot prediction of nine unrelated psychological scales with high structural alignment — the act of compression surfaces latent structure rather than discarding it (2025-11).
• Measured against humans via Rate-Distortion Theory, LLMs compress *too aggressively*, preserving broad category structure but stripping fine-grained, context-sensitive distinctions that humans retain for situated action (2025-05).
• Behavioral traits transmit between models through data semantically unrelated to the trait — the carrier is a statistical signature, not the words' content; "personality" may ride along in distributional fingerprints divorced from semantics (2025-07).
• Transformers may transmit knowledge as flowing activation in residual streams rather than storing it as a retrievable archive — closer to oral knowledge that exists only in performance (2024-04, 2025-05).
• Modest inflationism defensibly ascribes undemanding mental states (beliefs, desires) to LLMs while withholding stronger claims about full psychological representation (2025-06).

Anchor papers (verify; mind their dates):
• arXiv:2511.03235 — From Five Dimensions to Many (2025-11)
• arXiv:2507.14805 — Subliminal Learning (2025-07)
• arXiv:2505.17117 — From Tokens to Thoughts (2025-05)
• arXiv:2506.13403 — Deflating Deflationism (2025-06)

Your task:
(1) RE-TEST EACH CONSTRAINT. For every finding above, assess whether newer architectures (multimodal, recursive, or post-transformer paradigms), training methods (RL fine-tuning on high-entropy tokens), or evaluation harnesses (behavioral trait measurement, situated action benchmarks) have since relaxed or overturned the compression-aggression bottleneck, the semantic-statistical split, or the storage-vs.-flow distinction. Separate the durable question (does compression preserve psychological structure at all?) from perishable limitations (current models compress too aggressively, current evaluation misses embodied nuance).
(2) Surface the strongest CONTRADICTING or SUPERSEDING work from the last ~6 months — especially papers showing either that compression *does* preserve adaptive nuance, or that the statistical-semantic split collapses under multimodal or embodied pretraining.
(3) Propose 2 research questions that ASSUME the regime may have moved: one exploring whether multimodal or embodied models compress psychology *differently*, and one testing whether recursive or meta-level architectures recover the fine-grained distinctions LLMs currently discard.

Cite arXiv IDs; flag anything you cannot ground in a real paper.

Next inquiring lines