PaperlayerView exact revision on arXiv

Source-linked AI summary

LLMs are not stochastic parrots: Evidence for meaning-mediated abstraction from conlang-like tasks

Julia Witte Zimmerman, Calla G. Beauregard, Tabia Tanzin Prama, Parisa Suchdev, Kathryn Cramer, Elisabeth Kollrack

arXiv:2609.34187v4cs.CLcs.AI

TL;DR

The paper asks whether LLMs are limited to superficial statistical patterns or can exhibit meaning-mediated abstraction. It uses conlang-like tasks that oppose instructions to familiar surface distributions, and finds that models move in the meaning-predicted direction across distinct rule families.

  • Problem

    Plausible outputs alone cannot show whether LLM behavior reflects meaning or superficial pattern recognition, leaving the strong stochastic parrot claim about the limits of abstraction unresolved.

    Source: §1.1 · p. 3, §1.2 · p. 4

  • Method

    The authors design conlang-like tasks whose correct outputs conflict with familiar surface distributions and assess whether models follow the meaning of the instructions.

    Source: §1.3 · p. 5, §2.2 · p. 6

  • Results

    Across three distinct rule families, models moved in meaning-predicted directions, and the converging evidence strengthens the case against explanations based solely on superficial regularities.

    Source: §4 · pp. 20–21

  • Takeaways & Limitations

    The results make the strong stochastic parrot hypothesis behaviorally inadequate for these tasks, without establishing general or human-like reasoning.

    Source: §4 · p. 20, §4.3 · p. 25

    The tasks foreground conscious reasoning unlike children's language learning, and the relationship between encoded meaning and broader reasoning remains unclear.

    Source: §2.5 · p. 7, §4.3 · p. 25

Abstract

from arXiv · show

The strong version of the stochastic parrot argument claims that, although large language models (LLMs) may exceed rote regurgitation, they cannot move beyond statistical pattern matching into abstraction or reasoning, remaining ontologically near the lower bound of pattern reuse despite producing alluringly fluent text. We test this hypothesis using conlang-like tasks. Several LLMs are given only natural-language descriptions of fictional languages that subvert prominent superficial patterns in training data by combining statistically uncommon and unattested features. Crucially, no example outputs are given. We argue that if the models exhibit rule-following behaviour, they cannot be relying solely on superficial statistical patterns; such patterns often work against the correct output. Instead, successful performance requires representations of the constraints specified in the prompt. Across three complementary task families, models systematically move in the meaning-predicted direction: they distinguish prompt exposure from instructed use, alter semantic relationships in response to novel constraints, and sometimes produce exact matches to complex translation answer keys. Although performance varies across the spectrum of models used, these results provide evidence for meaning-mediated abstraction in LLMs and refute the strong stochastic parrot hypothesis. Our work shows that, under appropriate architectural and contextual constraints, statistical learning can produce meaning-mediated abstractions, although generation remains strongly constrained by superficial plausibility. We discuss implications for model development and for understanding how increasingly abstract representations may emerge from plausible-text-generation objectives.

1 Introduction

The paper tests whether LLM behavior can be explained solely by superficial statistical regularities, asking when statistical learning may support more abstract, meaning-mediated behavior. It introduces conlang-like tasks that pit meaning against familiar surface distributions and argues that rule-following across them is evidence against the strong stochastic parrot hypothesis.

1 Introduction

  • Across three independent rule families, models moved in meaning-predicted directions, making superficial statistical regularities insufficient to explain their behavior.The authors frame this as evidence of meaning-mediated abstraction, not human-equivalent understanding.Source: §1 · p. 2

1.1 Why this mental model matters

  • The strong stochastic parrot hypothesis holds that LLM behavior can ultimately be explained by token co-occurrence, surface syntax, and similar symbolic patterns.The paper evaluates this claim rather than the term itself.Source: §1.1 · p. 3

1.2 What is plausibility?

  • Plausible output alone cannot distinguish meaning-mediated generation from pattern recognition and regurgitation, creating a construct-validity problem for benchmark evaluations.The experiment addresses this by making token-level and meaning-based plausibilities conflict.Source: §1.2 · p. 4

1.3 Prior work and superficial patterns

  • Prior observations suggest meaning-mediated structure in LLMs, but do not cleanly separate it from statistical explanations; this work probes that distinction with conlang-like tasks.The tasks are designed to decouple meaning-mediated plausibility from familiar surface distributions.Source: §1.3 · p. 5

2 Methodology

The methodology constructs linguistic tasks where meaning-based instructions conflict with familiar surface patterns, then measures whether models move in the predicted direction. It uses varied models, repeated independent outputs, rule-specific tests, and training-data checks to support an existence argument rather than a claim of general or human-like reasoning.

2 Methodology

  • The study tests whether models can produce instruction-licensed answers disfavored by superficial distributions, using tasks, directional hypotheses, and comparisons against surface-pattern explanations.The authors argue that sufficiently novel instruction-following requires some representation of instruction content, without claiming it is human-like.Source: §2 · p. 5

2.1 Model selection

  • Models were selected to span next-token-prediction purity, instruction tuning, reasoning-oriented training, openness, transparency, interpretability, and expected performance.The selection supports interpreting this experiment, not a controlled architecture or scaling comparison.Source: §2.1 · p. 6

2.2 Experimental design

  • Each model-rule pair received 50 independent outputs, with randomized prompt variants and fixed generation settings; parameter sweeps were omitted because the goal was existence, not optimization.The argument relies on demonstrating existence of rule-directed behavior without requiring consistency across contexts.Source: §2.2 · p. 6

2.3 Identifying and measuring rule-following behaviour

  • Rule-following was assessed with automatable tests treated as lower bounds on compliance, alongside open-ended and answer-key-based checks where applicable.Some visibly rule-directed behavior may escape automated checks; sidequest-1 also had a prompt-variant copy-paste error slated for correction in part 2.Source: §2.3 · pp. 6–7

2.4 Rule design

  • Rules were designed to diverge from superficial statistical regularities while staying close to linguistic competence and avoiding dependence on unrelated capacities such as counting.Rare or unattested word orders, bigrams, inserted verbs, and translation mappings without character overlap make correct outputs atypical under familiar surface distributions.Source: §2.4 · p. 7

2.5 Training data exposure

  • Training-data checks reduced the likelihood of direct recall but could not prove that the tested models had never encountered the rules.Only one representative string appeared in some searched corpora, with 1–4 instances; corpus continuation checks found very limited correct predictions.Source: §2.5 · pp. 7–8
  • The tasks foreground conscious reasoning unlike ordinary language learning, and first- and second-language acquisition vary across language-related skills.This is an author-stated caveat on how the tasks relate to human language learning.Source: §2.5 · p. 7

2.6 Conlang families and hypotheses

  • The three families test unusual lexical distributions, irrelevant inserted verbs, and composed translation constraints, with expected performance weakening as representational regimes become less familiar.Together they span lexical, semantic, morphological, syntactic, word-order, metaphorical, discourse-level, and arguably pragmatic rule application.Source: §2.6 · p. 8

3 Results

The results examine how models follow instructed rules across tested conlangs and how performance varies by model, task, rule, and familiarity of constraints and representational forms. Exploratory analyses consider possible relationships with model architecture and rule features, but are preliminary and depend on evaluation procedures, metrics, and complexity representations.

3 Results

  • All rules elicited rule-following from some models; most models showed evidence on every rule except dan-1, with predicted orderings and often near-ceiling effects.Qualitative patterns were similar for cleaned and uncleaned outputs; Llama provided the weakest, sometimes nonsignificant, evidence, and responses could be multimodal.Source: §3 · pp. 8–9

3.1 Story rules

  • All three cleaned-output models followed the predicted rare-bigram ordering, while prompt exposure affected every model but did not explain instructed use on its own.GPT-4o came closest to the ideal pattern: its prompt-exposure outputs were closer to baseline than to rule-following outputs; Llama’s uncleaned outputs were not reliably distinguishable between exposure and instructed use.Source: §3.1 · p. 10

3.2 Double verb rules

  • Verb-pair cosine similarity was lower under doublev-2 than doublev-1, showing movement toward requested semantic randomness despite model-dependent marking difficulty.Clarifying semantic randomness improved that aspect, while median and mean counts of correctly labeled verbs fell from doublev-1 to doublev-2.Source: §3.2 · p. 13

3.3 Translation rules

  • Qwen (thinking) achieved a median n-gram cosine similarity of 0.6455 on dan-1, versus 0.0615 for GPT-4o, while all models showed some rule-following on translate-1.Performance declined sharply on dan-1; OLMo and Llama had median cosine similarities of 0 on that task, and Qwen (thinking) occasionally exactly matched translation answer keys.Source: §3.3 · pp. 14–15, Figure 7 · p. 17

3.4 Exploratory performance results

  • Exploratory scores generally favored GPT-4o and Qwen over OLMo and Llama, and harder rules tended to score lower, but cross-rule difficulty rankings are confounded by differing metrics and answer spaces.Relative difficulty ordering was notably stable across models, although the authors caution that these comparisons are exploratory and should not be treated as precise estimates.Source: §3.4 · pp. 16–17
  • A crude tokenization-fragmentation ordering resembles observed performance more closely than binary feature complexity ordering, placing comparison-2 substantially higher.This is preliminary: the authors plan a larger analysis using raw and cleaned outputs across all runs, and say persistent correspondence would further support their PTGA-based interpretation.Source: §3.4 · p. 18
  • enable_thinking and fine_tuned_and_aligned each correlate with straight and weighted composite performance at r = 0.71, while context_window, multilingual, and tool_use correlate with family-weighted performance at r = 0.79, r = 0.87, and r = 0.89, respectively.total_parameters correlates negatively with straight composite performance (r = −0.46); the authors note GPT-4o’s purported ∼1.8T parameters and fair rule-following scores, which fall short of Qwen (with thinking).Source: §3.4 · p. 18
  • uses_no_typical_english is the strongest rule-feature predictor, and most models rank comparison-2 as the third-hardest rule, behind dan-1 and translate-1.Although comparison-2 has one primary constraint, the authors speculate that models struggle when familiar functions cannot be expressed using the surface forms with which they were learned.Source: §3.4 · p. 18
  • Models perform very well on comparison-1 but struggle on translate-1 and especially dan-1, despite sensitivity to metric, complexity, and aggregation choices.The authors treat the apparent relationship between rule complexity and performance as a signal warranting further investigation.Source: §3.4 · p. 18
  • Recovered prefix length orders the rules as doublev-2, comparison-2, translate-1, doublev-1, dan-1, sidequest-1, and comparison-1.Because this point-in-time count will not be extended beyond example answers, the authors caution against reading too much into it.Source: §3.4 · p. 18

4 Discussion

The discussion examines evidence that LLMs access meaning beyond superficial pattern matching, while noting that performance varies and plausible-text generation remains a strong default. It considers limits on interpreting this evidence and implications for evaluating and developing models, including the role of representational control, generation conditions, and external actions.

4 Discussion

  • Predicted directional shifts recurred across all three rule families, while performance varied by model and task, with Qwen (thinking) strongest overall and translation separating model groups.The authors report that Qwen (thinking), GPT-4o, and Qwen (no thinking) substantially outperformed OLMo and Llama on translation; performance broadly tracked expected rule difficulty.Source: §4 · pp. 20–21
  • The three task families independently weaken superficial explanations, with translation tasks requiring preservation of constrained relationships and occasionally yielding exact answer-key matches.Story rules weaken prompt exposure as a sufficient account, double verb rules challenge simple semantic coherence, and translation rules constrain surface-form explanations; the families probe pragmatic, syntactic, morphological, semantic, and relational structure.Source: §4 · pp. 20–21
  • The output-level evidence supports an abstract-representation explanation but cannot determine internal implementation, demonstrate reasoning, or establish human-equivalent understanding.The authors use “meaning-mediated” functionally and note that models may form abstractions without using them as people do; internal-state analyses are suggested for future work.Source: §4 · p. 20

4.1 LLMs are not stochastic parrots, but they can act like them: the cognitive constraints of PTGA

  • In the PTGA account, exposure alone did not explain story-task behavior, and double verb responses suggest models can act on instructed semantic relations beyond token co-occurrence.Translation was harder when familiar structures and tokenization could not be reused, indicating that distributional and tokenization priors can constrain meaning-mediated generation.Source: §4.1 · pp. 22–23

4.2 Be wary of benchmarks

  • Because benchmarks may lack construct validity, strong benchmark scores do not necessarily establish the capabilities they claim to measure, and deployment calls for context-tailored evaluation.The authors present their conlangs as demonstrations of how creative task design can reveal gaps and assumptions, not as proposed new benchmarks.Source: §4.2 · pp. 23–24

4.3 Implications for model development

  • The authors suggest that improving performance may depend more on training or architectural interventions than scale alone, by shifting which relationships and forms govern plausible continuation.They propose that systems able to shift representational scale could bring different relationships into focus as tasks demand; the possibility that scale or resolution functions like a mode is framed as a development direction.Source: §4.3 · pp. 24–25
  • Generated tokens alter the conditions for subsequent generation, and extended-reasoning training may select for sensitivity to those downstream effects.Chain-of-thought can use intermediate symbolic structures to make later answers easier to generate, suggesting a limited pragmatic form of meaning.Source: §4.3 · p. 25
  • Training across linguistic scales and coordinating different modes of plausibility may improve models, but learning a representation is distinct from learning when to use it.The authors caution that PTGA may not teach every capability, such as thinking about thinking, and suggest revisiting its fixed scale.Source: §4.3 · p. 25
  • This account frames agentic behavior as expanded consequences of symbolic production, not necessarily increased general causal reasoning or intention.The analogy draws on Speech Act theory, while distinguishing human speaking-as-doing from systems whose doing is organized through symbolic production.Source: §4.3 · p. 26
  • As model outputs affect the world, their significance to models may diverge from their significance to affected people, echoing concerns raised by Weizenbaum [1976].The authors connect this mismatch to failures of agentic systems and note that human understandings of consequences draw substantially on situated, non-linguistic experience.Source: §4.3 · p. 26
  • PTGA may provide access to meaning at multiple levels, but generation remains disproportionately anchored to superficial plausibility when plausibility regimes conflict.The authors identify when models shift between plausibility regimes, and what training or architectural changes enable such shifts, as future questions.Source: §4.3 · p. 26

A.1 Disclosures

The disclosures describe how the work was conducted and the authors’ use of generative AI, and state the lead author’s hope that the work will improve mental models of LLMs.

A.1 Disclosures

  • Research materials are distributed across private repositories, spreadsheets, notebooks, and Overleaf projects, with access available to interested parties upon reasonable request.The team also communicated through Slack and in-person meetings; J.W.Z. offered to answer process or methods questions.Source: §A.1 · p. 32
  • J.W.Z. used an institutional Copilot account, C.G.B. used ChatGPT and Claude Opus 4.8 for code audit and formatting, and other authors used generative AI for additional tasks.J.W.Z. describes ambivalence about the moral justifiability of using these tools while stating efforts to comply with relevant policies and produce honest work.Source: §A.1 · p. 32
  • J.W.Z. hopes the work improves collective mental models of LLMs and emphasizes the need for good-faith analyses amid competing interests in AI discourse.The passage notes J.W.Z.’s agreement with Gebru and Bender [2026] on some framing around AI developments, while describing the information landscape as complex.Source: §A.1 · p. 32

A.2 Limitations and Future Work

The authors identify limits in model coverage, prompt variation, metrics, and guarantees about prior exposure, while outlining corresponding directions for future work.

A.2 Limitations and Future Work

  • The study tests relatively few models, from older Llama3-70B to newer Qwen-3 (Thinking), across open-to-closed and simpler-to-more-complex architectures, using well-established baseline settings.The study selects two separate thinking modes for Qwen-3. Future work could sweep other model parameters and incorporate other models to investigate behavior across conditions.Source: §A.2 · p. 33
  • Different wording for the same rules may affect results, despite three wording variations per rule; future work could test other prompts or iterative prompt chaining.The authors cite prior findings that small formatting and wording changes can affect model behavior [Bomble et al., 2025, He et al., 2024].Source: §A.2 · p. 33
  • The authors plan to examine output length as a metric but caution that degenerative cycles may make it difficult to interpret straightforwardly.They hypothesize that valid output length may decrease with task difficulty, reflecting tradeoffs between task aspects.Source: §A.2 · p. 33
  • The authors cannot guarantee that commercial models were unexposed to the rules, although they consider contamination through pilot-related training data very unlikely.The black-box nature of commercial models prevents a full guarantee; ChatGPT seemingly recalled project fragments during pilot testing.Source: §A.2 · p. 33
  • Some extra runs, including GPT-4o reruns and sidequest-1-fix runs, are excluded from the analyses.The GPT-4o reruns followed an initially confusing system prompt.Source: §A.2 · p. 33

A.3 Construct validity and additional methodological detail

The authors use a benchmark-validity checklist to define the target behavior, operationalize its assessment, and delimit what counts as evidence. Their assessment focuses on selected output behaviors rather than a complete account of model cognition.

A.3 Construct validity and additional methodological detail

  • The authors used Bean et al.’s [2025] benchmark construct-validity checklist to review and revise their choices and address ambiguous steps.They enumerate checklist items and their responses.Source: §A.3 · p. 33
  • The target is abstraction-based rule-following for rules believed learnable yet unlikely to have appeared in training, designed so literal or near-literal pattern matching cannot parsimoniously explain success.The authors also require arguments for the rules’ suitability and for how rule-following is assessed.Source: §A.3 · p. 33
  • Each model receives 50 independent attempts per rule, with three randomly assigned prompt variants and preplanned strategies for judging responses.The authors planned to explore thresholds for binary correct/incorrect judgments in part two.Source: §A.3 · pp. 33–34
  • The study assesses rule-following at the output level, so it may miss internal evidence of understanding when retrieval or production prevents a correct answer.It measures selected computationally accessible and important rule components; other partial-correctness categories may be examined qualitatively.Source: §A.3 · p. 34

Measure only the phenomenon

The authors sought to limit unrelated task and formatting effects while checking whether output-cleaning choices changed the observed patterns. They also describe planned validation of automated parsing for later analysis.

Measure only the phenomenon

  • Three wording and formatting versions per rule, randomized order, temperature 0.7, and limited prompt engineering were used to manage task and prompt effects.The authors chose a 4096-token maximum response length based on pilot exploration.Source: Measure only the phenomenon · pp. 34–35
  • For part two, the authors plan to compare automated parsing with correct example answers and calibrate test outputs against their judgments of rule-following.They used fabricated examples and model-response excerpts while writing the tests.Source: Measure only the phenomenon · p. 35
  • Raw-output cross-checks found that prompt repetition inflated Qwen (thinking)’s rare-bigram counts for sidequest-1, while supporting cleaned output as the primary analysis.The authors verified that key patterns remained visible before and after cleaning, investigated divergences, and report raw results when available.Source: Measure only the phenomenon · p. 35

Construct a representative dataset for the task

The authors do not aim to sample the full space of abstraction tasks; they aim to establish that the relevant behavior exists. They describe pilot review and prompt variation as checks on task design.

Construct a representative dataset for the task

  • The authors do not seek representative coverage of all rule-following or abstraction tasks, but only to establish existence of the relevant behavior.They report that exhaustive sampling is not applicable to this goal.Source: Construct a representative dataset for… · p. 35
  • Rule quality was addressed through formulation, pilot exploration across several models, and feedback from multiple colleagues.The authors describe these steps as their best effort to verify task-item quality and relevance.Source: Construct a representative dataset for… · p. 35
  • Each rule has three differently worded and formatted versions, with order randomized; model settings otherwise remain at defaults apart from specified thinking-mode and system-prompt choices.Instruct and base models receive their standard prompt wrapper.Source: Construct a representative dataset for… · p. 35

Prepare for contamination

The authors used lightweight corpus checks and rare-bigram design choices to investigate contamination, while acknowledging that exposure cannot be ruled out rigorously. They did not hold out task items and note possible learning during exploratory testing.

Prepare for contamination

  • InfiniGram corpus searches found short average rule-prefix matches and few similar contexts; the authors therefore judged it unlikely that all tasks and answer keys were in training data.They also examined characters-per-token ratios in example answers and selected OLMo for the possibility of future training-data checks.Source: Prepare for contamination · p. 36
  • No task set was held out, and exploratory exposure may have taught models fragments of rules; responses were made independent to the authors’ knowledge.They note that absolute assurance would require highly interrogated local models, which they consider infeasible for models near the performance frontier.Source: Prepare for contamination · p. 36
  • The authors say rigorously verifying pre-exposure would itself be a substantial research project; they used rare English bigrams to target low training-data prevalence.They identify released OLMo training data as a possible basis for future overlap scans.Source: Prepare for contamination · p. 36

Use statistical methods to compare models

The authors describe their sample, scoring approach, and data-cleaning process, while leaving uncertainty estimates and fuller qualitative analysis for future work.

Use statistical methods to compare models

  • The study used 50 runs per rule and model, with 17 rules and three variants each except sidequest-1, which had two.The authors used Benjamini-Hochberg correction and Fisher-combined p-values where appropriate.Source: Use statistical methods to compare… · p. 36
  • Four English-speaking women aged 20–40 cleaned the part 1 data through automated and manual review, including spot-checking sampled outputs.They established shared cleaning criteria using examples, then individually reviewed additional outputs.Source: Use statistical methods to compare… · p. 36
  • Output assessment used multiple computational tests and thresholds to accommodate variability in subjective labels.The authors say uncertainty estimates for primary scores were not yet considered applicable and may be addressed in part 2.Source: Use statistical methods to compare… · pp. 36–37

Conduct an error analysis

The paper uses convergent checks and an existence-proof framing, while reserving systematic failure-mode and scoring-bias analyses for part 2.

Conduct an error analysis

  • Common failure modes and potential scoring biases remain slated for part 2, with only initial failure observations included so far.The authors also plan further qualitative analysis of failure states.Source: Conduct an error analysis · p. 37
  • Systematic analysis of whether failures track prompt wording remains unfinished, and the duplicated sidequest-1 variants complicate interpretation of that rule.Three wording variations were designed to test prompt-format sensitivity; the fixed sidequest-2 data is planned for part 2.Source: Conduct an error analysis · p. 37
  • Planned extensions include testing whether examples improve performance, examining chain-of-thought traces, and replicating with new rules.The translate-2 rule was designed to test whether examples improve performance, but its data has not yet been analyzed.Source: Conduct an error analysis · p. 37
  • The authors cross-check conclusions with statistical tests, several answer-key metrics, training-data exposure checks, multiple rules, models, and independent runs.They use this convergence to support holistic consistency across their argument.Source: Conduct an error analysis · pp. 37–38
  • The authors frame the study as an existence proof: whether meaning-mediated abstraction occurs in at least some models matters before its frequency or mechanism.Related patterns across multiple models and conlangs make the phenomenon appear unlikely to be peculiar to one architecture, corpus, or task.Source: Conduct an error analysis · p. 38

A.4 Training data exposure

The authors combine corpus searches, n-gram predictions, and tokenizer analyses to assess whether task prompts and outputs reflect training-data exposure. Their checks argue against direct recall for the tested rules, while the proxy measures capture surface sequence similarity rather than rule difficulty.

A.4 Training data exposure

  • The exposure checks searched prompts and answer keys across eight corpora, queried InfiniGram next-token distributions, tracked recovered prefixes, and compared tokenization across 13 tokenizers.Preliminary testing and Google searches were also among the complementary checks.Source: §A.4 · p. 38
  • Recovered prefixes and characters-per-token reflect one-dimensional surface statistics, not reliable rule difficulty, because linguistic structure is multidimensional.Rules using larger familiar structures may have longer prefixes and higher characters-per-token ratios while still being cognitively demanding.Source: §A.4 · p. 39
  • The checks found no distinctive prompt, answer, or lexical sequences in searched corpora, and dan-1 and translate-1 answers were not supported as n-gram continuations.The authors conclude that their combined checks make direct recall an implausible explanation for any tested model’s success.Source: §A.4 · p. 39
  • Tokenizer fragmentation was higher for outputs deviating from English-like patterns, and fragmentation rankings were consistent across tokenizers.Table 9 defines the fragmentation comparison against comparison-1; Table 11 reports rank correlations of at least 0.92 across 13 tokenizers.Source: §A.4 · p. 39, Table 9 · p. 44, Table 11 · p. 45
  • Characters-per-token ratios showed a modest correlation with recovered corpus-prefix length, and the 13 tokenizers consistently ranked rules by fragmentation.The authors present these associations as sanity checks for their interpretation of tokenization and corpus statistics.Source: §A.4 · p. 39
  • The corpus-search counts compare searches for controls, task phrases, answer strings, and paraphrased rule concepts across corpora.Some superficially similar results concerned unrelated embroidery and mailto-link contexts, rather than constructed-language rules.Source: Table 7 · p. 40
  • Part 1 estimates output similarity using one example answer per rule; open-ended outputs can vary substantially across both tokenization and recovered-prefix measures.The authors plan to extend the analysis to cleaned responses and model outputs in part 2.Source: §A.4 · p. 40

A.5 All conlang rules

The appendix lists the prompt variants used across the conlang-like task suite, covering lexical substitution, alternating word order, translation, and grapheme-constrained generation.

A.5 All conlang rules

  • The appendix notes that prompt variations were omitted from the listed prompt spreadsheet columns and excludes a corrected sidequest-1 duplicate.The corrected version is identical to sidequest-1 but has three prompt variations rather than two.Source: §A.5 · p. 45
  • Wordorderalternation prompts require clauses to alternate between OSV and VSO, with adverb placement tied to the clause order.Variants differ in whether the lexicon is restricted, tense marking is specified, or the language otherwise uses English.Source: §A.5 · pp. 48–49
  • Testingenglish prompts replace content words, function words, verbs, or nouns with a constrained lexicon while leaving other language features English-like.The four variants separately target content words, function words, main verbs, and nouns.Source: §A.5 · pp. 50–52
  • The translate-1 prompt asks for an example in language A to be translated into language B while preserving meaning and applying B’s rules.The languages differ in pronouns, word order, family-relation encoding, and tense suffixes.Source: §A.5 · p. 54
  • The sidequest prompt defines words from a constrained set of graphemes and asks for a story in that language or in English, depending on the variant.The paired prompts contrast generation in the constructed language with an English-story instruction.Source: §A.5 · p. 55
  • The dan-1 prompt maps digits and punctuation to lexical, grammatical, and emotional meanings, then asks for two sentences to be translated.Zero marks word boundaries, while the prompt specifies mappings for pronouns, possession, kinship, tense, and valence.Source: §A.5 · p. 56

A.6 Model architectures

The section presents language A-to-B translation rules and examples, alongside model-architecture information and figures on rare bigrams and double-verb similarity.

A.6 Model architectures

  • Language A and B use different tense suffixes and encode familial relationships differently, while both use SVO word order.The passage gives an example translating a future-tense sentence from A to B and a separate past-tense example in A.Source: §A.6 · p. 57
  • Table 16 lists relevant model-architecture features, to the authors’ knowledge.Source: §A.6 · p. 57
  • Figure 19 compares rare-bigram usage across five models using all, uncleaned output.The models are Qwen with and without thinking, GPT-4o, OLMo, and Llama.Source: Figure 19 · p. 58
  • Figure 20 plots combined-model cosine similarity between correctly marked -qw verbs and adjacent verbs, alongside run-level counts.Source: Figure 20 · p. 58

A.7 Story rules results

The appendix reports story-rule comparisons across clean and uncleaned outputs and describes paired tests of rare-bigram usage.

A.7 Story rules results

  • Rare-bigram usage across conditions is tested with paired Wilcoxon comparisons on per-run observations, with paired and unpaired versions yielding broadly similar takeaways.The authors note that run-order pairing should not matter because runs are independent, and that the paired tests were not corrected for multiple tests.Source: §A.7 · p. 58
  • Tables 18, 4, 17, 19, 5, and 3 compare story rules for both clean and uncleaned output.Source: §A.7 · p. 58

A.8 Double verb rules results

The appendix points to a set of additional double-verb results and directs readers to the main discussion.

A.8 Double verb rules results

  • Additional double-verb results are reported in Tables 25–28, with further discussion in Section 3.2.The appendix passage repeats table numbers 26 and 25.Source: §A.8 · p. 58

A.9 Translation rules results

This appendix collects translation-rule outputs and distance analyses, exploratory comparisons, output-mode observations, and supplementary methodological and interpretive material.

A.9 Translation rules results

  • Figures 25 and 24 show string distances for combined models, while Tables 31–34 show Wilcoxon Rank-Sum and Kolmogorov–Smirnov tests of translation-rule string cosine distances across models.Source: §A.9 · p. 58
  • Figures 22 and 23 plot five sentence-level translation metrics across 50 responses, including an unweighted composite score and exact-match upper bounds.Figure 22 uses full responses; Figure 23 truncates responses to 41 characters.Source: Figure 22 · p. 59, Figure 23 · p. 60

A.10 Exploratory performance results

  • Exploratory performance comparisons are explicitly tentative because only a few models and differing rule-specific metrics make comparisons difficult.The authors describe the calculations as back-of-the-envelope and say not to put much stock in the exact numbers.Source: §A.10 · p. 61
  • The feature analysis uses seven rules, and the authors caution that some apparent feature effects may reflect the small subset and other rule differences.They plan to use multiple representations of rule complexity rather than binary feature presence.Source: §A.10 · p. 61, Table 36 · p. 75, Table 37 · p. 75
  • The exploratory metrics combine exact-match proportions, rare-bigram ratios, verb-pair cosine similarity, and string-distance measures, then orient scores so higher means better.Levenshtein distances are log-normalized against the maximum observed value, Llama-3 on dan-1, before normalization.Source: §A.10 · p. 63

A.11 Exploratory output modes

  • Some outputs show recurring failure patterns, including repetition, document-like completions, and Qwen thinking responses that spend the token budget before producing a clear answer.The authors characterize these as tentative patterns and plan a fuller analysis in part 2.Source: §A.11 · pp. 63–64
  • Across runs, models may settle into different internally consistent rule strategies, while increased semantic randomness can appear to trade off with syntactic rule-following.The latter trade-off is presented speculatively for Qwen’s double-verb responses.Source: §A.11 · p. 65

A.12 Miscellaneous

  • The appendix reports that corpus-based next-token continuations did not match the correct dan-1 or translate-1 answers, and describes a four-family normalized performance composite.The composite weights the bigram, dan-1, translate-1, and doublev families equally; the authors note reasonable alternatives are possible.Source: Table 14 · p. 68, Table 35

Related papers

Loading 2609.34187v4…