The emergence of flexible perspective reasoning in large language models
Bibliographic record
Abstract
Work on human reference processing has shown that, in sentences like “Mary asked her daughter Sally if she understood the assignment”, readers overwhelmingly interpret “she” as co-referring with “Sally”. This reflects perspective inference, or reasoning about who possesses at-issue information, and is inconsistent with a statistically-learned bias toward subject antecedent selections. The flexibility of inferencing is evident from the effect of manipulating the object character description (“Mary asked her tutor…”), where readers now prefer Mary as the antecedent. Until recently, these patterns have been largely unaccounted for by large language models (LLMs). Leveraging advancements in LLM interpretability techniques, the present study systematically examines how LLMs fare in relation to human judgments. We determine which layer activations impact these inferences and perturb them to causally link activations to model performance. Finally, we examine performance across training iterations, analyzing the point where subjecthood biases become evident and when more nuanced inferencing emerges.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.008 | 0.003 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.003 | 0.013 |
| Science and technology studies | 0.001 | 0.006 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.005 | 0.002 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.004 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; both teacher heads agree on what is shown here.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".