{"id":"6d4e6ac4-03fe-4ed2-a6dd-781fef40e4c3","arxiv_id":"2412.07646","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"LLM agents in a referential game gradually turn random artificial vocabularies into more structured and learnable languages, but iterated transmission also produces non-humanlike degenerate signals.","lead":"Two large language models played a guessing game with invented words, and the languages they used became more structured as the game went on. The authors then passed those languages to new model generations, which learned them more easily but also turned them into weirder, less human-like vocabularies.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The §3.2 vocabulary-update rule hard-codes alignment; without a no-update/no-communication control, TopSim gains cannot be attributed to LLM biases.","rationale":"The reader's weakest assumption identifies the load-bearing confound: the hard-coded vocabulary update in §3.2 is exactly the step that makes causal attribution to LLM biases insecure. My stress-test found no other objection that more directly threatens the central claim. The paper is internally consistent and honestly reports exploratory, non-humanlike outcomes, but the main result is underdetermined by the reported protocol because the update rule guarantees vocabulary alignment independently of any emergent communicative success. The requested controls, artifact release, and more conservative causal language would address the concern. Therefore the existing CONDITIONAL verdict remains appropriate, with no change needed from this pass.","tokens_in":15885,"tokens_out":4936,"duration_ms":51294,"concrete_test":"Run a matched control with the same 15 seeds and prompts but with the communication-block update rule disabled, so that after the labelling block each agent's stored vocabulary stays fixed while the agents still alternate generating and guessing for four rounds. Compute TopSim and PercCom on the same rounds. If TopSim still rises and PercCom still exceeds chance, the LLM's own in-context generalization drives the effect; if TopSim does not rise above the labelling-block level, the §3.2 update rule is the proximal cause of structure emergence. A second arm that overwrites both vocabularies with signals sampled from the existing vocabulary, rather than the speaker's generated signal, can further separate the update mechanism from the content of LLM-generated signals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that LLM biases shape unstructured languages toward structure depends on the communication block's update rule, which \"updat[es] the current stimulus to be associated with the produced signal\" (§3.2). This rule overwrites both agents' stored mappings with the speaker's produced signal on every interaction, so convergence to a shared vocabulary is enforced by the algorithm rather than demonstrated as an emergent outcome of communication. The paper includes no control condition that omits this update, omits communication, or replaces the LLM-produced signals with a null or random generator. Consequently, the observed TopSim increase, Ngram decrease, and roughly 75% PercCom could be explained by the hard-coded teaching signal plus in-context prompt structure, rather than by an implicit structural bias of the LLMs. The later iterated-learning chains inherit this confound because the transmitted vocabularies already contain update-induced regularities. This does not prove the claim false; it means the experiment as run cannot distinguish the contributions of the LLM's in-context generalization from the contributions of the hand-coded alignment mechanism.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a series of simulations in which two Llama 3 70B agents play a referential game over an initially holistic artificial language, with iterated-learning chains in which one agent's vocabulary is transmitted to the next generation. The authors measure communicative success (PercCom), topographic similarity (TopSim), Ngram diversity, and generalization (GenScore), and report that communication increases structure, structured languages generalize better, and generational transmission increases learnability while producing degenerate vocabularies. The central claim is that LLM implicit biases shape unstructured languages toward structure, making LLMs useful model learners for language evolution research.","tokens_in":16097,"tokens_out":10286,"duration_ms":95730,"significance":"If the central attribution is established, the paper would be a valuable demonstration that instruction-tuned LLMs exhibit inductive biases in artificial-language formation, extending prior work by Galke et al. (2023) and Kirby et al. (2015) and providing a tractable computational substrate for iterated-learning experiments. The study uses standard metrics, transparent statistical models with conditional and marginal R2, and reports quantitative results with uncertainties; it also explicitly acknowledges prompt sensitivity, single-model, and small-N limitations. However, the current design does not separate the hard-coded vocabulary-update mechanism from LLM-generated structure, so the headline claim is not yet supported. The specific chance-level error in Section 5.2 is also directly relevant to the conclusion that communication is successful.","major_comments":[{"comment":"The communication block hard-codes a vocabulary-alignment step: \"updating the current stimulus to be associated with the produced signal\" after interactions. The paper reports no control condition that omits this update, omits communication, or replaces the LLM-produced signals with a random or static generator. Because this update overwrites both agents' stored mappings for a stimulus with the speaker's signal, it externally forces convergence and alters the prompt vocabulary for all later blocks. The observed TopSim increase, Ngram decrease, and roughly 75% PercCom in Sections 5.2-5.3 are therefore compatible with an explanation in which the update rule plus the structured prompt format, rather than any implicit structural bias of the LLMs, drives the results; the iterated-learning chains inherit the same confound. Please add control simulations to isolate the LLMs' contribution.","section":"Section 3.2"},{"comment":"The task is to discriminate \"a target stimulus from four distractor stimuli\", which is naturally read as five alternatives, so chance-level communicative success is 20%, not 25% as stated. The 70% and 75% PercCom figures would still be above chance, but the size of the effect is smaller and the statistical comparison to chance should be recomputed and reported with the correct baseline. If the design instead used four total alternatives, the wording should be corrected to avoid ambiguity.","section":"Section 5.2"},{"comment":"Only the vocabulary of the agent with the highest TopSim is transmitted to the next generation. This selection rule can itself produce the observed increase in learnability and decrease in Ngram diversity, because high-TopSim vocabularies are already more regular and easier to reproduce. A control that transmits a randomly chosen agent, or that averages over agents, is needed to support the claim that generational transmission per se amplifies structural properties.","section":"Section 6"},{"comment":"The claim that increases in structure \"benefit communicative success positively\" (beta = 0.035) is in apparent tension with the earlier statement that PercCom does not increase significantly across rounds. The regression model is not described fully enough to know whether this coefficient reflects within-round time dynamics, between-simulation variation, or a combination; please clarify the specification and report the relationship between PercCom and TopSim conditional on round.","section":"Sections 5.2-5.3"}],"minor_comments":[{"comment":"Specify whether the vocabulary update is applied after every interaction or after every round, and whether it applies to both agents' stored mappings; the current sentence is ambiguous and the magnitude of the confound depends on this timing.","section":"Section 3.2"},{"comment":"The statement that \"initial explorations with GPT-3.5 and Llama 2 7B showed similar behaviours\" is not accompanied by data; if this claim is retained, a supplementary table or figure should be provided.","section":"Section 3.3"},{"comment":"The explanation that LLMs produce longer signals because their context windows are \"considerably larger\" is asserted without a manipulation or analysis; present it as a hypothesis rather than as a conclusion.","section":"Section 7"},{"comment":"The paper uses \"learnability\" to refer to the edit distance between ground-truth and reproduced signals in the labelling block; this operationalization should be stated more explicitly to avoid conflating reproduction accuracy with acquisition in the human sense.","section":"Section 6"},{"comment":"There is a typo: the text refers to \"Figure 11\" when describing the four simulation blocks, but the corresponding figure in the manuscript is Figure 1.","section":"Section 3"},{"comment":"The paper does not provide a code or data availability statement; for a computational study of this type, a reproducibility statement would be helpful.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The update-rule confound is the decisive issue; I would advise the editor that a revision with control conditions (no update, no communication, random speaker) is necessary before the central claim can be accepted. The chance-level error and the transmission-selection issue are secondary but must also be addressed. I do not see grounds for rejection: the experimental framework and metrics are sound, and the required controls are feasible within the scope of the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: this paper is the first to put two LLM agents in a Kirby-style referential game and run iterated learning over invented vocabularies. That is a real step beyond Galke et al.'s use of GPT-3 on human-evolved languages. The core results—TopSim rises, Ngram diversity falls, higher TopSim predicts better generalization, and generational transmission improves learnability—are internally consistent and statistically solid for the design. I believe the trends are real.\n\nWhat the paper does well: the setup is a close adaptation of Kirby et al. (2015) and Raviv et al. (2021), with a sensible use of linear mixed-effects models, and the authors are honest about several limitations (single model, greedy decoding, prompt sensitivity). The qualitative example in Table 1 is compelling: the LLM does compose sub-word units for shape and amount. If you want to see an LLM building a mini-language, this is a nice demonstration.\n\nThe soft spot is the one the stress-test flags. In §3.2, the communication block updates 'the current stimulus to be associated with the produced signal.' That means every interaction overwrites both agents' stored mapping for that stimulus with the speaker's output. Convergence to a shared vocabulary is thus written into the algorithm; it is not an emergent property of communication. Without a control that omits the update, or replaces LLM-generated signals with random ones, the observed TopSim increase and the 75% PercCom cannot be attributed to the LLM's implicit structural biases. The iterated-learning chains inherit this confound because the transmitted vocabularies already contain update-induced regularities. This is a load-bearing gap, but it is fixable with the right controls.\n\nMinor cautions: only 15 simulations and six chains, a single model family, and no code or data release. Those are less concerning individually but add up.\n\nWho is this for? Anyone working on LLMs as models of language evolution or on emergent communication. It deserves a serious referee, but the referee should demand the controls and a revised claim that structure emerges from the LLM-plus-update-rule system, not from the LLM alone.","headline":"Novel LLM referential-game study with a hard-coded alignment rule that muddies the attribution of emergent structure to LLM biases.","tokens_in":16647,"tokens_out":3434,"would_cite":true,"duration_ms":30695,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"LLM agents shape unstructured artificial languages into structured ones through communication.","keywords":["language evolution","emergent communication","large language models","referential game","iterated learning","compositional structure","topographic similarity","artificial language learning"],"falsifier":"Run the same four-block simulation with the communication block removed (agents only label and re-label the same stimuli) or with the stimulus–signal update rule replaced by a random association; if TopSim still rises significantly and n-gram diversity still falls, the claimed causal role of communicative interaction and LLM bias would be falsified. A more targeted test would be to hold the language fixed and swap in a model with a much smaller context window; if the degenerate long-signal behaviour disappears, memory constraints rather than bias explain it.","tokens_in":15655,"feed_emoji":"🗣️","tokens_out":7380,"duration_ms":58415,"temperature":0.7,"pith_summary":"This paper asks whether large language models, when they must learn and use an artificial language in a referential game, impose the same kind of structural biases that human learners do. The authors show that two LLM agents communicating about colored shapes gradually reshape an initially random holistic vocabulary so that similar meanings get similar signals, a property known as topographic structure. This structure improves the agents' ability to generalize to unseen meanings, and passing the language down to new generations of LLM learners makes it easier to learn. However, the same cultural-transmission process can drive the vocabulary toward degenerate, underspecified forms that look non-humanlike, such as longer signals and many-to-one mappings. The central claim is that LLMs are viable model learners for language-evolution experiments, not just for evaluating pre-existing structure.","feed_headline":"LLM agents shape unstructured languages into structured ones","feed_subtitle":"Two chatbot agents converge on compositional signals, but transmission can degrade them into non-humanlike forms.","key_machinery":"The Lewis referential game played by two LLM agents, in which the speaker produces a signal for a stimulus, the listener picks the most probable prefilled distractor, and after each interaction the current stimulus is associated with the produced signal in the next prompt. The game is embedded in a four-block pipeline (guessing, labelling, communication over four rounds, testing), and the iterated-learning version transmits the vocabulary of the highest-TopSim agent to the next generation. Topographic similarity, the Mantel-test z-score between signal edit distances and semantic distances, is the primary structure metric, paired with n-gram diversity and a generalisation score for unseen stimuli.","core_discovery":"The paper's central discovery is that initially unstructured holistic languages are indeed shaped to have structural properties when optimised for the implicit biases of LLM agents. In the authors' simulation, two instruction-tuned LLMs play a Lewis referential game: one produces a signal for a target stimulus, the other guesses the target among distractors, and the interaction updates each agent's vocabulary by associating the produced signal with the current stimulus. Over four rounds, topographic similarity (TopSim) rises significantly and n-gram diversity falls, which the authors interpret as the emergence of compositional structure. The structured vocabularies then support better generalisation to unseen stimuli, measured by a generalisation score. Iterated learning across eight generations reduces the edit distance between ground-truth and produced signals, meaning the languages become easier for LLMs to learn, but it also yields degenerate vocabularies (repeated signals for different meanings) and longer signals, behaviours that diverge from human experimental findings. The authors take this as evidence that LLMs can serve as flexible tools in simulations of language evolution, and that the set-up exposes the models' own structural preferences.","pith_inferences":["The built-in vocabulary-update rule, which associates each produced signal with the current stimulus after every interaction, may be driving most of the convergence rather than the LLMs' communicative reasoning; without a control condition that removes communication or randomizes the association, the causal attribution to LLM biases remains open.","The absence of a genuine memory bottleneck in LLMs, which have large context windows, could explain why iterated learning produces degenerate underspecified vocabularies instead of the compressed compositional ones seen in human experiments.","The paper's metric suite could be ported to other instruction-tuned models or to decoding strategies with higher temperature; if structure emergence disappears under those variations, the effect is tied to the specific model's biases rather than a general property of LLMs.","A direct human-LLM comparison in the same referential game would be the natural next test of whether LLMs truly serve as model learners in the sense claimed; the authors note such experiments are future work."],"forward_implications":["LLM-based agents can substitute for human participants in at least some language-evolution experiments, lowering cost and enabling large-scale systematic manipulations.","The finding that structured languages support better generalization in LLMs suggests that compositional pressures benefit machine learners, not just human learners.","The observed non-humanlike drift (longer signals, degenerate vocabularies) implies that LLM-driven iterated learning must be interpreted with caution; the models' memory and decoding biases are not human-like.","Because communicative success rose only modestly and varied across simulations, the setup is sensitive to initial conditions and may need explicit expressivity pressures to maintain human-like structure.","The increase in learnability across generations demonstrates that weak inductive biases in LLMs can be amplified by cultural transmission, a process previously documented in humans."],"supporting_citations":[{"why":"Supplies the reference-game and iterated-learning design, and the human experimental results that the LLM outcomes are compared against.","marker":"Kirby et al. (2015)"},{"why":"Provides the GenScore metric and the three-attribute meaning space used to test generalisation.","marker":"Raviv et al. (2021)"},{"why":"Showed that compositional structure is advantageous for GPT-3 in learning artificial languages, providing the prompt-format basis and the prior evidence that structure aids LLMs.","marker":"Galke et al. (2023)"},{"why":"Demonstrates that language models struggle to learn unnatural word orders and non-hierarchical structure, motivating the claim of a structure bias in LLMs.","marker":"Kallini et al. (2024)"},{"why":"Establishes the iterated-learning paradigm and the method for generating initial random CV-syllable signals.","marker":"Kirby et al. (2008)"},{"why":"Provides the Mantel test used to compute topographic similarity from signal and meaning distance matrices.","marker":"Mantel (1967)"},{"why":"Supplies the interaction-based vocabulary update rule that the communication block relies on.","marker":"de Boer (2000)"},{"why":"Argues that expressivity pressure prevents degenerate vocabularies, the expectation against which the paper's observed underspecification is measured.","marker":"Smith et al. (2013)"}],"fun_headline_variants":["LLM agents shape unstructured languages into structured ones","Chatbots evolve structured signals, but transmission degenerates them","LLM referential game yields compositional languages","Structure emerges in LLM communication, then degenerates over time","LLMs impose structure on artificial languages, but learning degrades"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim depends on the assumption that the observed structure emerges from the LLMs' own inductive biases, yet the setup includes a hard-coded vocabulary-alignment step that updates the current stimulus to be associated with the produced signal, and no control condition omits communication or this update rule.","fun_headline_variants_meta":{"raw":{"variants":["LLM agents shape unstructured languages into structured ones","Chatbots evolve structured signals, but transmission degenerates them","LLM referential game yields compositional languages","Structure emerges in LLM communication, then degenerates over time","LLMs impose structure on artificial languages, but learning degrades"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1326,"prompt_tokens":906,"completion_tokens":420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":522,"completion_tokens_details":{"reasoning_tokens":342}},"tokens_in":522,"tokens_out":420,"duration_ms":4816,"temperature":1.0,"reasoning_tokens":342,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:38:11.301994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four-block simulation with the communication block removed (agents only label and re-label the same stimuli) or with the stimulus–signal update rule replaced by a random association; if TopSim still rises significantly and n-gram diversity still falls, the claimed causal role of communicative interaction and LLM bias would be falsified. A more targeted test would be to hold the language fixed and swap in a model with a much smaller context window; if the degenerate long-signal behaviour disappears, memory constraints rather than bias explain it.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Mantel test used to compute topographic similarity from signal and meaning distance matrices."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the interaction-based vocabulary update rule that the communication block relies on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Argues that expressivity pressure prevents degenerate vocabularies, the expectation against which the paper's observed underspecification is measured."}],"review_version":1}