First-hand information for everyone
Main discovery: Bilingual language models (GPT-2 variants) capture aspects of cross-linguistic influence. They explain additional variance in non-native (L2) English reading times beyond a monolingual English model. This L1-specific advantage is strongest when the model’s L1 matches the reader’s L1, and it concentrates in later eye-tracking measures. (d1, d2)
How the study tested this: The authors trained GPT-2 bilingual models for each L1–English pair. They used a sequential bilingual protocol with four Steps of Exposure (SoE) where English training was introduced at steps 0, 16k, 32k, or 48k. All models ran for 64k training steps. A monolingual English baseline used the same tokenizer. Training data came from FineWeb splits. Sentences with code-switching were filtered. A SentencePiece tokenizer was trained per L1–L2 pair on balanced data. From each language corpus the authors sampled one billion tokens for LM training. (m1, m2)
Evaluation and metrics: Models were evaluated against the Multilingual Eye Movement Corpus (MECO L2). The corpus contains eye-tracking recordings of participants from multiple L1 backgrounds reading the same 12 English texts. The study compared model surprisal to four word-level reading-time measures: first fixation duration, gaze duration, go-past time, and total fixation duration. Surprisal is the negative log-probability the model assigns to a word (summing subword surprisals when words are tokenized into pieces). Predictions were tested with linear mixed-effects models that include baseline covariates (word frequency, length, position, and preceding-word features) and crossed random intercepts for participants and words. The bilingual vs. monolingual gains were summarized with a pseudo-R^2 (∆R^2), and matched vs. mismatched comparisons used likelihood-ratio tests on nested models. (m3, m4)
Concrete results: Matched bilingual models (model L1 = reader L1) outperform the monolingual baseline and mismatched bilingual models for L2 readers. This matched advantage is concentrated in later processing measures: gaze duration, go-past time, and total fixation duration. For the earliest measure, first fixation duration, bilingual surprisal gives negligible or no predictive advantage. The native English control group stays near zero, indicating the advantage is specific to L2 readers. Nested model comparisons show that, except for first fixation, matched models contribute more unique variance and have higher mean ∆LL and higher significance ratios than mismatched models. Go-past time shows a clear upward trend in matched-model advantage as SoE increases. (c1, c2, d2)
Script overlap and timing: Writing-system overlap between L1 and English modulates when the bilingual advantage appears. Latin-script L1s show a bilingual advantage already in gaze duration. Non-Latin-script L1s typically show little or no early advantage, but both script groups show comparable or superior bilingual performance in later measures. Non-Latin groups show a steady upward trend in go-past time advantage as SoE increases. (c3)
Proficiency effects: Reader proficiency (measured primarily with LexTALE) modulates the bilingual advantage nonlinearly. For late measures (go-past time and total fixation duration) the bilingual ∆R^2 follows an inverted-U across LexTALE. The advantage peaks for intermediate readers (roughly 65–80% LexTALE) and decreases for the most advanced readers (>90% LexTALE). A PCA-derived composite proficiency index replicates this inverted-U pattern for late measures. (c4)
Significance: These results provide evidence that bilingual LMs encode L1-specific predictive biases that align with human cross-linguistic influence during L2 reading. The effects are most visible in measures linked to later-stage processing, such as integration and reanalysis. The stepwise exposure manipulation (SoE) and the matched-versus-mismatched comparisons support an L1-specific source for the explained variance. (d1, d2, m1, m3)
Limitations and caveats: The study evaluated only English as the L2 and used a single eye-tracking corpus (MECO), so generalization to other target languages or corpora is not established. Some MECO L1 subsets were recorded in regions with regional varieties (e.g., Brazilian Portuguese) while model pretraining used standard corpora; this mismatch could add noise. Script overlap was examined but other typological factors (word order, morphology) were not independently isolated. Proficiency was operationalized mainly via LexTALE, a lexical measure; higher-level comprehension or syntactic awareness were not directly measured. All model configurations used a single initialization seed, so variability across random initializations was not estimated. Finally, code-switching was filtered with sentence- and word-level detectors, but fully removing code-switching remains an open challenge and residual instances could affect cross-linguistic transfer. These limitations qualify how broadly the findings can be applied. (l1, l2, l3, l4, l5, l6)
Takeaway: In MECO English reading, bilingual GPT-2 models encode L1-specific biases that improve prediction of late-stage eye-tracking measures, with effects sensitive to script overlap, exposure timing, and learner proficiency. (d1, d2, c1, c2, c3, c4, l1, l4)