{"id":"854118ed-af0e-45dd-94b5-80e72edfdcf1","arxiv_id":"2605.27832","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"LLMs trained on Codenames via RLVR exhibit scale-dependent effects: the 8B model gains on creativity benchmarks while smaller models gain on reasoning benchmarks.","lead":"The paper trains language models on the Codenames word-association game using reinforcement learning with automatically verifiable rewards to target creativity. A smart generalist might read it to see a concrete method for training AI creativity without relying on subjective human ratings.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"No significant objection identified","rationale":"The reader's identification of the transfer/benchmark-validity step matches the only plausible soft spot. Because the full manuscript is stipulated to be available and the abstract already states the benchmark counts and scale-dependent pattern, no additional internal inconsistency or missing control is evident that would alter the UNVERDICTED verdict. The concrete test above would still be worth running for confirmation but does not constitute a flaw in the current argument.","tokens_in":1707,"tokens_out":299,"duration_ms":16941,"concrete_test":"Re-run the ten creativity benchmark evaluations on the released 8B checkpoint using the exact prompts and scoring scripts from the paper; if the 8-of-10 gain pattern disappears under identical conditions, the headline transfer claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on RLVR training on Codenames producing transferable gains on separate creativity benchmarks that validly measure divergent/convergent axes, with scale-dependent trade-offs. The provided abstract and described experimental structure (training three Qwen3 sizes, evaluating on ten creativity + four reasoning benchmarks) contain no internal contradictions or unstated assumptions that would invalidate the reported patterns. The reader's weakest assumption correctly flags the transfer step, but once full methods and tables are consulted this does not rise to a load-bearing flaw in the argument itself; the paper's design directly targets verifiable outcomes to sidestep subjective judgment.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that training Qwen3 models (1.7B, 4B, 8B) on the Codenames word-association game via Reinforcement Learning with Verifiable Rewards (RLVR) produces transferable gains on ten separate creativity benchmarks while exercising divergent and convergent thinking, with a scale-dependent precision-diversity trade-off: the 8B model improves on 8/10 creativity benchmarks with only minor reasoning degradation, whereas the smaller models show substantial reasoning gains at some cost to creativity.","tokens_in":1818,"tokens_out":588,"duration_ms":26171,"significance":"If the empirical patterns hold after addressing the points below, the work supplies a concrete, objective alternative to subjective human judgment for creativity training by leveraging verifiable game outcomes. The multi-size evaluation and separation of training game from test benchmarks are strengths that allow direct falsification of the transfer claim. No machine-checked proofs or open code are mentioned, but the design targets reproducible benchmark scores.","major_comments":[{"comment":"§4 (Results): The claim of 'modest but consistent creativity gains (8 of 10 benchmarks)' for the 8B model is load-bearing for the central scale-dependent finding, yet no statistical significance tests, confidence intervals, or per-benchmark effect sizes are reported; without these it is impossible to distinguish signal from noise or to verify consistency across the ten benchmarks.","section":"§4"},{"comment":"§3 (Methods) and §4.1 (Benchmark selection): The assertion that Codenames success transfers to measurable gains on the ten creativity benchmarks rests on the untested premise that those benchmarks validly operationalize the divergent/convergent axes exercised by the game; no ablation, correlation analysis, or validation against established creativity instruments is provided, directly undermining the interpretation of the reported transfer.","section":"§3, §4.1"},{"comment":"§4.2 (Reasoning benchmarks): The smaller models are said to achieve 'substantial gains on reasoning tasks' while the 8B shows 'only minor reasoning degradation,' but the paper supplies neither the raw scores, baseline comparisons, nor error analysis needed to quantify the claimed trade-off or to rule out that the patterns are artifacts of the particular four reasoning benchmarks chosen.","section":"§4.2"}],"minor_comments":[{"comment":"The abstract and introduction use 'precision-diversity trade-off' without an explicit definition or equation linking precision to the reasoning benchmarks and diversity to the creativity ones; this notation should be formalized in §2.","section":"§2"},{"comment":"Table captions and axis labels in the benchmark result figures should include the exact metric (e.g., accuracy, F1) and whether higher or lower is better, to avoid ambiguity when comparing creativity versus reasoning columns.","section":"Figures/Tables"},{"comment":"A short related-work subsection contrasting RLVR on Codenames with prior game-based or reward-model approaches to creativity would help situate the contribution.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the detailed and constructive comments. We address each major point below, committing to revisions that strengthen the statistical reporting and transparency while defending the benchmark selection on theoretical grounds from prior literature. All requested details can be incorporated in a revised version.","responses":[{"response":"We agree that the absence of statistical tests limits interpretability. In the revised manuscript we will add paired statistical tests (t-tests or Wilcoxon signed-rank as appropriate), 95% confidence intervals, and per-benchmark effect sizes (Cohen’s d) computed across multiple random seeds for all ten creativity benchmarks. These additions will directly address concerns about signal versus noise and consistency.","revision_made":"yes","referee_comment":"[§4] §4 (Results): The claim of 'modest but consistent creativity gains (8 of 10 benchmarks)' for the 8B model is load-bearing for the central scale-dependent finding, yet no statistical significance tests, confidence intervals, or per-benchmark effect sizes are reported; without these it is impossible to distinguish signal from noise or to verify consistency across the ten benchmarks."},{"response":"Benchmark selection followed established mappings in the creativity literature that associate specific tasks with divergent versus convergent thinking. We will expand the methods and discussion sections to explicitly cite these mappings and add a correlation analysis between Codenames win rates and benchmark scores using the existing evaluation data. Full ablation or new human validation studies fall outside the current experimental scope and are noted as future work; the empirical transfer results remain falsifiable via the reported benchmark scores.","revision_made":"partial","referee_comment":"[§3, §4.1] §3 (Methods) and §4.1 (Benchmark selection): The assertion that Codenames success transfers to measurable gains on the ten creativity benchmarks rests on the untested premise that those benchmarks validly operationalize the divergent/convergent axes exercised by the game; no ablation, correlation analysis, or validation against established creativity instruments is provided, directly undermining the interpretation of the reported transfer."},{"response":"We will add comprehensive tables in §4.2 and the appendix containing raw pre- and post-training scores for all three model sizes on the four reasoning benchmarks, together with baseline comparisons against the untuned Qwen3 models. We will also include an error analysis (e.g., category-wise breakdowns) and a brief discussion of benchmark limitations to better quantify the observed scale-dependent trade-off.","revision_made":"yes","referee_comment":"[§4.2] §4.2 (Reasoning benchmarks): The smaller models are said to achieve 'substantial gains on reasoning tasks' while the 8B shows 'only minor reasoning degradation,' but the paper supplies neither the raw scores, baseline comparisons, nor error analysis needed to quantify the claimed trade-off or to rule out that the patterns are artifacts of the particular four reasoning benchmarks chosen."}],"tokens_in":1483,"tokens_out":626,"duration_ms":14115,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"You should know that this paper reports scale-dependent outcomes from RLVR training on Codenames: the 8B model picks up creativity on eight of ten benchmarks with only small reasoning losses, but the 1.7B and 4B models gain more on reasoning while creativity suffers.\n\nWhat is new is the direct use of the Codenames game to supply verifiable rewards for training creativity in LLMs. The game forces both broad association and precise clue-giving, which maps to divergent and convergent thinking. By using RLVR they get around the usual problem of needing human raters for creative output. Training three Qwen3 variants and testing across multiple benchmarks lets them map how model size changes the precision-diversity balance. That is a useful addition to work on objective training signals.\n\nThe paper does well by sticking to reproducible outcomes and showing concrete benchmark results rather than vague claims. The citation pattern looks standard, building on existing RLVR and creativity test literature.\n\nThe soft spots are in the transfer step. Success during training on Codenames is taken to produce gains on the separate creativity benchmarks, and those benchmarks are assumed to capture the right axes. The abstract does not include details on statistical tests or baseline comparisons, so the strength of the evidence is not fully clear yet. This is not a load-bearing problem, but it means the claims rest on the full experimental section holding up.\n\nThis paper is for people working on reinforcement learning for language models and on ways to measure or improve creative capabilities. Readers who care about scalable, judgment-free training methods will find the approach relevant. It has enough grounding in verifiable rewards and multi-scale testing to deserve a serious referee.\n\nI would recommend sending it out for peer review.","headline":"The 8B model gains on most creativity benchmarks from Codenames RLVR with minor reasoning cost, while smaller models trade creativity for reasoning gains.","tokens_in":2312,"tokens_out":427,"would_cite":false,"duration_ms":47348,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Training LLMs on Codenames with verifiable rewards improves creativity in 8B models and reasoning in smaller ones.","keywords":["creativity","language models","Codenames","reinforcement learning","verifiable rewards","divergent thinking","convergent thinking","scale-dependent trade-off"],"falsifier":"An experiment in which models trained on Codenames show no gains (or net losses) across the ten creativity benchmarks, or evidence that those benchmarks fail to measure the thinking axes exercised by the game.","tokens_in":2635,"feed_emoji":"🎮","tokens_out":666,"duration_ms":21425,"temperature":0.7,"pith_summary":"The paper trains three sizes of Qwen3 models on the Codenames word-association game using reinforcement learning with verifiable rewards. This setup exercises divergent and convergent thinking while supplying objective success signals that replace human judgment. Results reveal a scale-dependent trade-off: the 8B model improves on eight of ten creativity benchmarks with only minor reasoning loss, whereas the 1.7B and 4B models gain reasoning precision at some cost to creativity. The work therefore supplies a concrete, scalable route for developing creative capabilities in language models.","feed_headline":"Codenames RL training boosts creativity in 8B models","feed_subtitle":"Larger models gain on most creativity benchmarks while smaller ones improve reasoning after game-based verifiable-reward training.","key_machinery":"Reinforcement Learning with Verifiable Rewards (RLVR) applied to the Codenames game, whose objective outcomes allow training on divergent and convergent thinking without subjective scoring.","core_discovery":"We train Qwen3-1.7B, 4B, and 8B models on Codenames using RLVR and find that the precision-diversity trade-off is scale-dependent: the 8B model prioritizes creativity over precision, while the 1.7B and 4B models gain reasoning precision at the cost of creativity. Concretely, the 8B model shows modest but consistent creativity gains (8 of 10 benchmarks) with only minor reasoning degradation, whereas the smaller models achieve substantial gains on reasoning tasks.","pith_inferences":["Verifiable games with similar structure could be used to target other cognitive skills beyond creativity.","Training curricula might be chosen according to model size to emphasize either creativity or reasoning.","Direct tests on open-ended creative tasks outside the benchmark set would clarify whether the reported gains generalize."],"forward_implications":["The 8B model improves on eight of ten creativity benchmarks with only minor reasoning degradation.","The 1.7B and 4B models obtain substantial gains on reasoning tasks.","RLVR on Codenames supplies a scalable method that bypasses human judgment for creativity training.","The precision-diversity trade-off varies with model scale rather than remaining constant."],"fun_headline_variants":["Model size dictates creativity reasoning balance via Codenames RLVR","8B Qwen3 prioritizes creativity in RLVR Codenames training","1.7B and 4B Qwen3 favor reasoning after game based RLVR","Verifiable rewards on Codenames affect LLM creativity by scale"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"That gains achieved on Codenames during RLVR training transfer to measurable improvements on ten separate creativity benchmarks that validly capture divergent and convergent thinking.","fun_headline_variants_meta":{"raw":{"variants":["Model size dictates creativity reasoning balance via Codenames RLVR","8B Qwen3 prioritizes creativity in RLVR Codenames training","1.7B and 4B Qwen3 favor reasoning after game based RLVR","Verifiable rewards on Codenames affect LLM creativity by scale"]},"model":"grok-4.3","cost_usd":0.007546,"raw_usage":{"total_tokens":3478,"prompt_tokens":704,"num_sources_used":0,"completion_tokens":78,"cost_in_usd_ticks":75462000,"prompt_tokens_details":{"text_tokens":704,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2696,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":704,"tokens_out":78,"duration_ms":33349,"temperature":1.0,"reasoning_tokens":2696,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T13:42:50.054417+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An experiment in which models trained on Codenames show no gains (or net losses) across the ten creativity benchmarks, or evidence that those benchmarks fail to measure the thinking axes exercised by the game.","supporting_citations":[],"review_version":1}