{"id":"477cabe8-4975-489e-9217-d10dafd53ab5","arxiv_id":"2411.15386","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Fine-tuning BERT-family language models on moral reasoning or fMRI data does not significantly improve how closely their internal activations match human brain activity during moral judgment, although larger models align better.","lead":"This paper measures how closely the internal representations of language models match human brain activity during moral reasoning, using fMRI data from people judging moral scenarios. It finds that fine-tuning the models on ethics or brain data does not reliably increase that match, and that larger models tend to match better.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BrainScore's sensitivity is the load-bearing link; Appendix B's own admission that vision-control scores were not beaten suggests the Table 2 null may be a false negative.","rationale":"The reader's weakest assumption was measurement sensitivity; this is indeed the load-bearing point, and the manuscript itself supplies evidence for it in Appendix B. I considered other candidate concerns — absence of significance tests for Table 2, post hoc selection of the LAST sampling method, and possible data leakage between fine-tuning and BrainScore evaluation — but the vision-control admission is more central because it directly tests whether the metric isolates moral reasoning. If the BrainScore cannot beat a control region expected to be irrelevant, then the null result in Table 2 is weak evidence against the effectiveness of fine-tuning: it may simply mean the measurement is too noisy or too dominated by generic activity. The paper has real strengths: many fine-tuning runs, a public code link, and a self-critical appendix that explicitly reports the failure of the control group. My concern is not an accusation of misconduct; it is a validity concern about a null result whose interpretation depends on the sensitivity of the metric. The conditional verdict is appropriate: the paper should be accepted only if the positive-control sensitivity check supports the metric, or the conclusions should be softened to 'we could not detect an improvement' rather than implying fine-tuning does not work. The proposed test is computationally cheap and directly settles whether the null result is informative.","tokens_in":7721,"tokens_out":6617,"duration_ms":63414,"concrete_test":"Take the exact Section 3.4 pipeline and the held-out fMRI data used for Table 2, using, e.g., RoBERTa-large LAST. Inject a known moral-reasoning-specific signal by scaling the predicted activations in the moral/ToM NeuroSynth ROIs by factors of 1.05, 1.10, and 1.20 of the ROI standard deviation, while holding vision ROIs fixed. Recompute BrainScore for each scaling. If BrainScore does not increase monotonically with the injected signal or the increase is smaller than the noise floor implied by the Table 2 differences, the metric lacks the sensitivity needed to support the claim that fine-tuning produced no improvement. If it does increase cleanly, the null result is credible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 4 — that fine-tuning fails to improve BrainScore — is a null result whose force depends entirely on the sensitivity and construct validity of the BrainScore metric in Section 3.4. Appendix B contains the paper's own admission: 'our models were not able to achieve better brain scores than the control group, meaning that our experiment did not achieve the desired effect.' Since vision was chosen as a control that was expected not to improve, this admission indicates that the measured alignment is not specific to moral reasoning / Theory of Mind. If the regression of [CLS] activations onto 1,024 DiFuMo parcels and the subsequent correlation with held-out fMRI is dominated by generic stimulus-driven or language-related variance, the small differences in Table 2 (e.g., BERT-large 0.217 vs 0.213; DeBERTa 0.271 vs 0.273) are within noise and would not be expected to reflect a real representational change. The paper reports no confidence intervals or significance tests for the BrainScore differences in Table 2, so 'unable to significantly improve' is not statistically established. The concern is not that the authors are wrong about what they measured; it is that what they measured may be too blunt an instrument to detect the improvement they set out to test, making the headline negative result weak evidence for 'the importance of gathering more data.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript asks whether fine-tuning encoder language models (BERT-base/large, RoBERTa-large, DeBERTa-v2-xlarge) on moral reasoning behavioral data (ETHICS), fMRI data (Koster-Hale et al. 2013), or both, improves their BrainScore, defined as the correlation between model activations and held-out fMRI activity fitted to 1,024 DiFuMo parcels. The authors report two main findings: (1) fine-tuning can improve ETHICS Commonsense accuracy, sometimes beyond the original benchmark report, but adding fMRI data does not consistently beat ETHICS-only fine-tuning; and (2) BrainScores do not significantly improve over pre-trained models after any fine-tuning protocol, while larger models tend to have higher BrainScores. The paper concludes that the negative result supports the need for larger moral reasoning datasets.","tokens_in":7887,"tokens_out":5003,"duration_ms":43171,"significance":"The paper is, to my knowledge, the first to measure and attempt to improve BrainScore in the moral reasoning domain, and the scale of the fine-tuning effort (450 runs, four models) is a strength. If the null result were rigorously established, it would be a useful cautionary data point for the NeuroAI literature and for claims that human brain data can straightforwardly induce human-like representations in LLMs. The authors also release their code and are unusually explicit about the failure of their intervention. The main limitation is that the null result is only as informative as the BrainScore measurement, and the manuscript does not currently demonstrate that this measurement is sensitive to the targeted moral/ToM dimension or that the observed differences are statistically meaningful.","major_comments":[{"comment":"The central negative claim that fine-tuning does not significantly improve BrainScore is not supported by inferential statistics. The table reports only means and standard deviations, and the standard deviations (0.09–0.12) are as large as or larger than the differences being interpreted (e.g., DeBERTa 0.271 vs 0.273 and 0.237; RoBERTa 0.173 vs 0.144–0.156). I recommend reporting per-run paired comparisons between fine-tuned and pre-trained models, confidence intervals on the differences, and, ideally, a noise ceiling for the BrainScore metric so that a null result can be distinguished from an insensitive measurement.","section":"Section 4 / Table 2"},{"comment":"The interpretation of the null result depends on the sensitivity of BrainScore to the moral-reasoning/ToM dimension, and the paper's own Appendix B states that 'our models were not able to achieve better brain scores than the control group, meaniing that our experiment did not achieve the desired effect.' Since vision was chosen as a control expected not to improve, this admission suggests that the measured alignment may be dominated by generic stimulus- or language-related variance rather than by moral reasoning. Please provide evidence that the metric is sensitive to the targeted dimension, for example a positive control where a known alignment-boosting manipulation increases BrainScore, or a demonstration that BrainScore tracks ETHICS accuracy across models.","section":"Section 3.4 / Appendix B"},{"comment":"The BrainScore pipeline is described only qualitatively, which makes the headline null result difficult to evaluate or reproduce from the text. The manuscript should specify the regression inputs (which layers and tokens enter the regressor), the train/test split (subject-level or item-level, number of examples), the number of fMRI time points and the hemodynamic lag values, the cross-validation scheme, and the exact formula for aggregating layer-wise scores. Without these details, it is hard to rule out overfitting of the 1,024-parcel regression or misspecification of the temporal alignment as explanations for the null result.","section":"Section 3.4"},{"comment":"The claim that 'we could not improve accuracy by fine-tuning on the fMRI data only or on a combination of fMRI and ETHICS, compared to fine-tuning purely on ETHICS' is based on small numbers of runs with wide, overlapping confidence intervals (e.g., RoBERTa-large has 4 runs for ETHICS-only and 18 for the combined condition, with CS test set intervals of 53.0–91.4 and 49.7–91.6). This does not establish equivalence; a paired or Bayesian comparison across matched runs is needed before concluding that fMRI fine-tuning has no effect on ETHICS accuracy.","section":"Section 4 / Table 1"}],"minor_comments":[{"comment":"The text refers to 'Appendix A' for the vision control group, but Appendix A is the Author Contributions section; the relevant material appears in Appendix B. Please correct the cross-reference.","section":"Section 3.2"},{"comment":"There is a typo in the sentence 'our models were not able to achieve better brain scores than the control group, meaniing that our experiment did not achieve the desired effect'; 'meaniing' should be 'meaning'.","section":"Appendix B"},{"comment":"The abbreviation 'A VG' in the Sampling column is not defined in the table caption; Section 4 defines 'VG' as an average of all time points, but the table should be self-contained.","section":"Table 3"},{"comment":"The figure captions mention one-tailed Bonferroni-corrected significance, but the text never states which test was used, how many comparisons were corrected for, or whether the comparisons are across models, layers, or subjects. Please add this information.","section":"Appendix B (Figures 13–20)"},{"comment":"The column header 'Brain Score Mean Brain Score St. Dev.' is awkwardly worded and ambiguous; splitting it into 'Mean Brain Score' and 'Standard Deviation' would improve readability.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a timely negative result, and the authors are transparent about the null outcome and release code. However, the central claim is a null result whose force depends entirely on statistical power and metric sensitivity; the current manuscript lacks significance tests and contains an explicit admission that the desired effect was not achieved. I think the paper is potentially salvageable with a revised analysis that includes paired comparisons, a noise ceiling, and a positive-control or sensitivity analysis for BrainScore, but without those additions the headline conclusion is not yet established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the headline: this is the first BrainScore measurement for moral reasoning, and the main finding is a null result — fine-tuning BERT, RoBERTa, and DeBERTa on ETHICS, on moral-reasoning fMRI data, or on both does not improve BrainScore over the pretrained baselines. The novelty is real but local: the method is an established BrainScore pipeline applied to a new domain, and the result is a negative finding rather than a new mechanism.\n\nWhat the paper does well: it runs a large number of fine-tuning runs (450 runs, 1,082 models), releases code, and is candid about the limits of its own experiment. The ETHICS accuracy results in Table 1 show the fine-tuning pipeline works as intended (several numbers beat the original ETHICS paper), so the negative BrainScore result is not because the models failed to learn the task. There is no circularity: BrainScore is computed on held-out fMRI after fitting a regression on training data, and fine-tuning uses separate splits. The paper also reports layer-wise and region-specific scores in the appendix, which is more than most alignment papers do.\n\nThe soft spots are real. The central claim in Section 4 — 'unable to significantly improve our brain-score metric' — is not backed by any significance test or confidence interval for the BrainScore differences in Table 2. Looking at the numbers, e.g., BERT-large 0.217 vs 0.213, DeBERTa 0.271 vs 0.273, these differences look like noise, but the paper doesn't actually show that. The stress-test note about sensitivity is on point: if the regression onto 1,024 DiFuMo parcels and the subsequent correlation with held-out fMRI is dominated by generic language or stimulus variance, the null could be a false negative. Appendix B's own admission that 'our models were not able to achieve better brain scores than the control group' supports this worry — the vision control was expected to not improve, and the fact that it wasn't beaten suggests the metric isn't specific to moral reasoning. The post hoc selection of the LAST sampling method, based on accuracy, is another soft spot; it's a multiple-comparison issue that isn't addressed.\n\nNone of this is fatal. The paper is an honest, careful negative result, and the first in its domain. The force of the null claim depends on a metric whose sensitivity is not established. The authors should either add permutation tests or confidence intervals to show the differences are within noise, or report the vision control quantitatively and discuss why the metric is sensitive enough to detect real changes. As is, the paper is a useful preliminary report but the headline claim is under-supported.\n\nWho should read it: researchers working on brain-model alignment or NeuroAI, and anyone thinking about whether fine-tuning on neural data can steer LLM representations. I'd bring it to a reading group, but I wouldn't cite it as strong evidence until the statistical question is tightened. It deserves a serious referee — it's the first measurement in a new domain and an honest negative result — but I'd recommend conditional acceptance with requests for the missing statistics and a direct treatment of the sensitivity concern.","headline":"First BrainScore measurement for moral reasoning reports a null fine-tuning effect, but the null claim is under-supported without sensitivity or significance tests.","tokens_in":8476,"tokens_out":3105,"would_cite":false,"duration_ms":27249,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fine-tuning BERT, RoBERTa, and DeBERTa on moral-reasoning data, fMRI data, or both does not raise their BrainScore against human fMRI activity; larger models are more brain-aligned to begin with.","keywords":["brain-model alignment","BrainScore","moral reasoning","fMRI","fine-tuning","language models","ETHICS benchmark","representational similarity"],"falsifier":"Re-running the LAST-sampling fine-tuning protocol on a moral-reasoning fMRI dataset with, say, ten times more held-out trials, and observing a BrainScore improvement over the pretrained baseline that exceeds the standard deviations in Table 2, would overturn the paper's central claim; a positive-control experiment that recovers a known injected alignment signal would rule out measurement insensitivity.","tokens_in":7483,"feed_emoji":"🧠","tokens_out":10664,"duration_ms":81244,"temperature":0.7,"pith_summary":"The paper asks whether a language model's internal representations can be pushed closer to human brain activity in the moral domain by fine-tuning on human moral judgments, human fMRI responses, or both. It reports what it identifies as the first BrainScore measurement of moral-reasoning alignment for encoder-based language models. The central finding is a null result: none of the fine-tuning protocols raised BrainScore significantly above the pretrained baseline, while larger pretrained models already scored higher. Fine-tuning on the ETHICS commonsense benchmark improved task accuracy, but adding fMRI data did not improve accuracy further. If the null result holds, representational alignment with human moral cognition is not bought by current fine-tuning on small behavioral or neuroimaging datasets.","feed_headline":"No brain-score gain from moral-reasoning fine-tuning","feed_subtitle":"A first test with BERT, RoBERTa, and DeBERTa finds scale, not task training, drives fMRI alignment.","key_machinery":"The load-bearing instrument is the BrainScore metric: a Pearson correlation between predicted and actual fMRI activity, where predictions come from a fitted regression that maps [CLS] token activations of a language model to 1,024 regions of interest in the DiFuMo functional atlas. The fMRI stimulus set is a published moral-judgment dataset, preprocessed through a self-supervised learning pipeline, with four temporal sampling strategies (average of all time points, LAST, MIDDLE, SENTENCES) and four region groups (Theory of Mind, moral, language, vision) used as target and control partitions. The metric's job is to detect whether fine-tuning shifts the model's internal geometry toward the geometry of human moral reasoning; the paper's conclusion is that, at this resolution, it does not.","core_discovery":"On the authors' own terms, the discovery is a measured absence: fine-tuning language models on moral-reasoning data does not make their [CLS]-token representations more predictive of human fMRI activity. Using a BrainScore protocol that regresses model activations onto 1,024 DiFuMo atlas parcels and correlates predictions with held-out fMRI, the authors find no significant improvement for any model or sampling scheme. BERT-large-cased goes from 0.217 to 0.213, RoBERTa-large from 0.173 to between 0.144 and 0.156, and DeBERTa-v2-xlarge from 0.271 to between 0.237 and 0.273, all within one standard deviation of the baseline. The same fine-tuning does raise ETHICS commonsense accuracy, sometimes above the scores reported for the original benchmark, so the null is specific to neural alignment rather than to learning in general. The paper is also the first to attempt this measurement in the moral-reasoning domain.","pith_inferences":["Inference: the flat BrainScore may reflect that the [CLS]-head readout used for prediction discards the layer-wise features where alignment could change; probing individual attention heads or intermediate layers might reveal fine-tuning effects the aggregate score masks.","Inference: the result is consistent with the idea that representational convergence with brains is primarily a pretraining-scale phenomenon; a decisive test would be to compare BrainScore across model families at matched parameter counts before and after much larger-scale moral fine-tuning.","Inference: a promising next experiment is to fine-tune on naturalistic moral narratives or on a contrastive objective that aligns intermediate layers directly to fMRI parcels, rather than on classification labels, and to check whether BrainScore rises on held-out subjects.","Inference: because the held-out fMRI set is small and the regression is fit per atlas parcel, the null could be a false negative; re-running the same protocol on a moral fMRI dataset with more subjects or trials would be the most direct way to test the robustness of the conclusion."],"forward_implications":["Model scale, not task-specific fine-tuning, is the main driver of brain alignment in moral reasoning: the largest pretrained model's BrainScore exceeds the fine-tuned scores of smaller models.","Adding fMRI data to ETHICS fine-tuning does not improve ETHICS accuracy beyond ETHICS-only fine-tuning, so neural data is not an obvious free lunch for task performance.","BrainScore and benchmark accuracy decouple: a model can learn the ETHICS task while its representational alignment with fMRI stays flat.","The null result is stable across fMRI sampling methods (average, LAST, MIDDLE, SENTENCES), making it unlikely to be an artifact of a single time-point choice.","Researchers pursuing brain-aligned moral AI should expect that current fine-tuning recipes and dataset sizes are insufficient, and should treat moral-reasoning neural alignment as an open problem."],"supporting_citations":[{"why":"It supplies the moral-judgment fMRI dataset used as the alignment target.","marker":"[Koster-Hale et al., 2013]"},{"why":"It supplies the ETHICS commonsense benchmark and the accuracy baselines the paper compares against.","marker":"[Hendrycks et al., 2020]"},{"why":"It provides the preprocessed fMRI version already fit to the DiFuMo atlas.","marker":"[Thomas et al., 2023]"},{"why":"It defines the 1,024-region DiFuMo functional atlas that fixes the prediction target space.","marker":"[Dadi et al., 2020]"},{"why":"It establishes the BrainScore comparison methodology this paper adapts to moral reasoning.","marker":"[Schrimpf et al., 2020]"},{"why":"It is the prior positive result against which this paper's null result is compared.","marker":"[Aw and Toneva, 2023]"},{"why":"It is the earlier demonstration that fMRI-targeted fine-tuning can induce brain-relevant bias, which motivates the fine-tuning protocol.","marker":"[Schwartz et al., 2019]"},{"why":"It defines the BERT architecture used for the base and large models.","marker":"[Devlin et al., 2019]"},{"why":"It defines the RoBERTa-large architecture tested.","marker":"[Liu et al., 2019]"},{"why":"It defines the DeBERTa-v2-xlarge architecture tested.","marker":"[He et al., 2021]"}],"fun_headline_variants":["Fine-tuning for morality? Brains say no","Scale, not tuning, drives LLM-brain alignment","Moral fine-tuning: no BrainScore gain","Ethics training won't align LLMs with fMRI","LLM moral training: fMRI alignment unchanged"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The BrainScore measurement assumes that regressing [CLS] token activations onto 1,024 DiFuMo atlas parcels and correlating predictions with held-out fMRI activity is sensitive enough to detect a genuine improvement in representational alignment; if the regression fit is noisy, the held-out set is small, or the hemodynamic lag is mis-specified, the observed null result could be a false negative rather than evidence that fine-tuning cannot work.","fun_headline_variants_meta":{"raw":{"variants":["Fine-tuning for morality? Brains say no","Scale, not tuning, drives LLM-brain alignment","Moral fine-tuning: no BrainScore gain","Ethics training won't align LLMs with fMRI","LLM moral training: fMRI alignment unchanged"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000326,"raw_usage":{"total_tokens":1798,"prompt_tokens":892,"completion_tokens":906,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":834}},"tokens_in":508,"tokens_out":906,"duration_ms":7267,"temperature":1.0,"reasoning_tokens":834,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:20:57.674013+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the LAST-sampling fine-tuning protocol on a moral-reasoning fMRI dataset with, say, ten times more held-out trials, and observing a BrainScore improvement over the pretrained baseline that exceeds the standard deviations in Table 2, would overturn the paper's central claim; a positive-control experiment that recovers a known injected alignment signal would rule out measurement insensitivity.","supporting_citations":[],"review_version":1}