Umwelt Engineering: Designing the Cognitive Worlds of Linguistic Agents
Rodney Jehu-Appiah
cs.CL, cs.AI
2026-03-29
No-Have lifts ethical accuracy 19.1 pp and classification 6.5 pp in 4,470 trials; no constrained debugger beats 88.2% alone, but a 3-agent ensemble covers 100%.
Prompt engineering tunes what an agent is asked. Context engineering tunes what it knows at inference time. Both treat language as a transparent pipe: change the payload, leave the medium alone.
That assumption is easier to defend for people than for models. A human can finish a thought in spatial intuition, affect, or motor memory and only later put words on it. A standard LLM has no such remainder. Its reasoning happens in the token stream. The vocabulary it can use, the grammar it can instantiate, and the distinctions those patterns make speakable are the cognitive operations it can run.
Jakob von Uexküll called the slice of the world an organism can perceive its Umwelt. This paper parks the term on the agent stack and argues for a third layer, Umwelt engineering: deliberate design of the linguistic environment the model thinks in. Adjacent evidence already exists. Chinese-trained and English-trained models attend to causes differently. Plain text versus Markdown versus JSON can swing the same task by as much as 40%. Mentalese, a synthetic reasoning language, cuts tokens 4 to 16 times while keeping roughly 90 to 98% of accuracy. The claim under test is narrower and harder: freeze the items, change only the medium of reasoning, and see whether scores move.
Experiment 1 tests two vocabulary bans.
Three cheap instruction models: Claude Haiku 4.5, GPT-4o-mini, Gemini 2.5 Flash Lite. Seven task types, 130 items: syllogisms, causal reasoning, analogy, classification, epistemic calibration, ethical dilemmas, math word problems. Each item ran under three conditions, three models, and four repetitions (one at temperature 0, three at 0.7). Planned 4,680 trials, 4,470 completed, 4,344 scoreable.
Experiment 2 changes the question. Sixteen agents, each with a distinct linguistic constraint, debug 17 software problems (51 ground-truth findings) at temperature 0. The control gets a plain English debugging instruction. The metric is union coverage, not whether any constrained agent beats the control alone.
No-Have is the stabler intervention.
| Task | Control | No-Have | E-Prime |
| Ethical dilemmas | 76.6% | 95.6% (+19.1 pp) | 92.1% (+15.5 pp) |
| Classification | 93.0% | 99.6% (+6.5 pp) | 96.2% (+3.1 pp) |
| Epistemic calibration | 68.7% | 76.1% (+7.4 pp) | 63.0% (−5.7 pp) |
| Causal reasoning | 76.7% | 81.5% (+4.9 pp) | 90.8% (+14.1 pp) |
| Syllogisms | 100% | 97.9% (−2.1 pp) | 96.6% (−3.4 pp) |
Overall accuracy: control 83.5%, No-Have 88.6%, E-Prime 85.4%. No-Have compliance is 92.8%. E-Prime compliance is 48.1%, so half of those trials mixed banned copulas back in.
Cross-model gaps are sharper. Gemini's ethical baseline is 41.7%; No-Have adds 46.3 pp and E-Prime 42.3 pp, from a floor that inflates raw points. GPT-4o-mini under E-Prime drops epistemic calibration from 53.8% to 26.2% (−27.5 pp), consistent with losing copula hedges such as "this claim is well-supported." Haiku versus GPT-4o-mini E-Prime task deltas correlate at r = −0.75: the same ban reshapes cognition in opposite directions. Word count falls 16 to 33% on every non-math task.
In Experiment 2, no constrained agent beats the control's 88.2%. The 16-agent union covers 100%. A greedy 3-agent subset (analogical, counterfactual, minimal) also hits 100% at 17.6% of the full ensemble's API calls. Of 560 three-agent combinations, 8% reach full coverage, and every successful subset includes the counterfactual agent. It was the only one to surface a specification ambiguity: whether "preserving order" means first occurrence or last.
For people who ship agents, this is closer to changing the default lens than to writing a longer instruction. A cheap No-Have system prompt produces repeatable gains on ethics and classification, and models can actually keep the rule. E-Prime is a better probe than a knob: it exposes how different native Umwelten interact with the same ban, and it is a poor universal booster.
The multi-agent lesson is harder. Sixteen copies of one mode lose to three orthogonal linguistic modes. The counterfactual constraint did not raise solo accuracy, yet it contributed the finding nobody else saw. Teams already stacking ensembles can treat "what the agent is allowed to say" as a diversity axis next to "which model."
This is incremental evidence, not a new architecture. The constraints arrive as system prompts, so the border with prompt engineering is a spectrum. The paper says as much: the three-layer stack may collapse to a continuum, but the far end of that continuum is barely touched by current prompt practice.
The stated primary limit is the missing active control: a prompt as elaborate as E-Prime or No-Have that does not ban any words. Both constraint prompts are much longer than the control, so metalinguistic self-monitoring could be doing part of the work. The two bans still produce divergent task profiles (a 13.1 pp swing on epistemic calibration), which is a weak match for a pure length effect, and it still does not close the hole.
E-Prime's 51.9% violation rate mixes compliant and non-compliant reasoning. Gemini's low baselines (ethical 41.7%, causal 57.8%) inflate raw gains; syllogisms sit at a 100% control ceiling. All three models are cheap instruction followers, and Experiment 1 is multiple choice. Experiment 2 uses an LLM judge for semantic matching. The item banks are small: 130 items, 17 debug problems. Cross-model correlations rest on n = 7 tasks; r = −0.75 is suggestive.
The taxonomy lists eight traditions and the experiments test two. Candidate traditions were surfaced with Claude and then filtered by the author. Before treating this as a third layer in the stack, the program still needs that active control, plus a repeat on frontier models and open-ended generation.
Related work already cites Korzybski, Bohm, Bridgman, and Láadan. What is new is using vocabulary bans as LLM reasoning interventions with a three-model, seven-task crossover, not the invention of those linguistic traditions.