This work extends a recently introduced concept-circuit extraction method to a 2x2 design -- Python and Rust crossed with Qwen2.5-Coder-7B and DeepSeek-Coder-V1-6.7B -- and measures a complete inventory of grammatical concepts identically in all four cells.
Abstract
Do independently trained language models come to represent the same thing in the same way? We answer for code, extending a recently introduced concept-circuit extraction method to a 2x2 design -- Python and Rust crossed with Qwen2.5-Coder-7B and DeepSeek-Coder-V1-6.7B -- and measuring a complete inventory of grammatical concepts (58 Python, 57 Rust) identically in all four cells: the smallest design that separates what depends on the task, the language, and the model. The answer splits into three parts. What earns dedicated circuitry is set by the task: the models agree on which concepts receive circuits (Spearman $\rho$ = 0.638 for Python, 0.673 for Rust, both p<$10^{-7}$). Where those circuits sit is set by the model: Qwen processes concepts in a late band (~L17-19), DeepSeek at L6-7, for both languages. How circuits grow across layers is also set by the model: Qwen gives its atomic concepts an early spike that DeepSeek does not."Are circuits universal?"thus has no single answer: yes for What, no for Where and How -- universality is a property of representational content, not of computational organisation. None of this structure was fixed in advance. The agreement could have landed anywhere between independence and identity; it lands at $\rho \approx 0.65$. Rust constructs receive 2-3x more concept-specific circuitry than their Python equivalents, in both models. Both models share neurons between the languages (6/7 and 7/7 paired constructs), DeepSeek 1.94x more than Qwen -- a direction no prior result predicts. And Qwen binds nine keywords of Rust's type-and-trait machinery into one tight neuron cluster (Jaccard 0.535 vs null 0.112, p<0.001), a semantic dimension invisible in surface syntax. Ablation and linear probes confirm the circuits are functional. All claims are scoped to this 2x2; whether the per-model profile predicts a third model is the designed next test.
When a language model follows an in-context conditional rule such as"if P(x) then A else B,"does it assemble a runtime circuit with one module that tests the predicate and another that routes the answer? We probe this with activation patching under a four-donor design whose two swapped-rule donors make the condition and the answer word disagree, so each layer reveals which of the two it carries. Across three open models from two families and six languages sharing one fixed item bank, a mid-stack residual band carries the predicate's truth value: patching it reroutes the answer with predicate-outcome flip near 1.0 and mapping flip near 0.0, meeting a strict pre-specified isolation criterion in 17 of 18 cells, and the same localization holds across five predicate families. The router shows the opposite profile. A learned subspace flips A and B near-perfectly within the trained pair yet transfers to a new pair at approximately 0 in every model, while in Gemma-3-4B (the only model probed cross-lingually) it transfers at approximately 0.98 to the same pair in other languages. Under every probe we ran, the router direction is token-bound and non-transferable (largely answer-readout in Gemma, pair-specific in Qwen) rather than an abstract routing module. Test is modular; under these probes, route is not.
This work presents the first systematic study of inverse relation directionality in LLMs, using a benchmark consisting of 5,457 instances spanning 27 distinct inverse relation labels and reveals systematic asymmetries in inverse relation classification across LLMs.
Take three frontier mixture-of-experts models (Alibaba, OpenAI, NVIDIA; 3.6-4.0B active parameters each) and fine-tune them to reason in a low-resource language. On accuracy benchmarks almost nothing happens, and the benchmark itself is noise at this scale: changing only the random seed moves the score by 7.7 points, more than every data and recipe effect we measured. That null is our first result. The real changes live where accuracy cannot see. Base models never think in Greek: 0 of 1,000 reasoning traces, even when the question is Greek, so the model answers correctly while reasoning in a form its user cannot read, audit, or correct. After supervised fine-tuning (SFT), every released checkpoint reasons in the language of the question on ~98% of items, one family at 3x fewer tokens, with judged grammaticality improving on all four models and general ability within a few points of each base: nothing was forgotten, and fluency was gained. We propose six behavioural dimensions that make such changes measurable, each gated to reject any metric that correlates with output length, and we report how our own instruments lied: six failures, each caught by a control. What SFT cannot do is fix its own defects: a quarter of answers skip the requested format, answers leak into the reasoning channel, and an explicit"think in English"is obeyed under half the time. Reinforcement learning with verifiable rewards, pre-registered before training, fixes the first two outright (fallback 24% to 2.5%, leak 3.5% to 0.0%, both against a flat random-reward control) and moves the third (+9.1pp), while the Greek reasoning habit survives an accuracy-only gradient untouched. We release five checkpoints. The instruments, the controls and the pre-registration travel to any low-resource language; Greek is the case that let us measure them.
Large language models access knowledge inconsistently across languages, but to what extent do they differ in their skill sets when interacting with different languages? This work quantifies cross-lingual skill inconsistency orthogonally from knowledge and general benchmark performance. We do this via multilingual self-play: two instances of the same model compete in a text-based game, each interacting through a different language interface. Since the model, opponent, rules, state space, and available actions remain fixed, this setting isolates the effect of language on the model's realized behavior. We build a multilingual extension to TextArena and evaluate three open-weight models across eight languages and six games covering spatial reasoning, imperfect information, resource allocation, and repeated interaction. We find that the same model can exhibit markedly different playing strength across languages, with systematic variation in win--loss margins, invalid actions, and strategic tendencies. Detailed analyses reveal language-specific failures in spatial reasoning, card-conditioned decisions, and optimal move selection. In some settings, changing only the intermediate reasoning language recovers much of the lost performance, suggesting that language can affect different stages of the decision process. These results show that skill discrepancies are a measurable major roadblock in the development of truly multilingual models. Better understanding these discrepancies can help us design models that perform more equitably across languages.
Bobby Cheng, Adam Gaber, Zhengzhe Liu et al.· 0 citations
Chinese neologisms exploit diverse and unique linguistic mechanisms, such as phonetic substitution (e.g., 886 for ``bye-bye'') and visual character decomposition that are rare in other languages. We introduce CNeo-Bench, a benchmark of 4,759 such neologisms with reference definitions, organized into five top-level categories and nine subcategories by the linguistic mechanism behind each expression. CNeo-Bench is paired with a two-tier evaluation framework that separates whether a model can describe a neologism from whether it can operate on its underlying mechanism. Evaluating 18 LLMs, we find that Chinese neologisms remain an open challenge; most models fall below 40\% on definition generation, and on several subcategories a systematic recognition-manipulation gap emerges: models describe neologisms correctly but, in source-form restoration tasks, substitute a semantic equivalent (paraphrase) for the source form rather than producing the source form itself. A few-shot analysis on 1,058 hard items shows that in-context examples can solve many difficult cases, but leave a noticeable portion of errors remaining, indicating challenges beyond prompting alone can address.
Kaiyan Zhao, Zhongtao Miao, Zheyong Xie et al.· 0 citations
Abstract Language models (LMs) are commonly criticized for not “understanding” language. However, many critiques focus on cognitive abilities that, in humans, are distinct from language processing. Here, we instead study a kind of understanding tightly linked to language: inferring “who did what to whom” (thematic roles) in a sentence. Does the central training objective of LMs—word prediction—result in sentence representations that capture thematic roles? In two experiments, we characterized sentence representations in four LMs that have been proposed as models of human language processing. The overall representational similarity of sentence pairs did not reflect whether they had the same agent/patient assignments or opposite agent/patient assignments. Furthermore, we found limited evidence that thematic role information was available in any subspace of hidden activations. However, some attention heads robustly captured thematic roles, independently of syntax. Therefore, LMs can extract thematic roles but this information influences their representations weakly.
Joseph M. Denning, Xiaohan Guo, Bryor Snefjella et al.· Annual Meeting of the Cognit...· 1 citation
