{"id":"d04cbaff-38f4-4b09-a1f8-532e67b17eba","arxiv_id":"2608.07208","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A linear probe on frozen LLM activations comes within 0.6 to 2.1 accuracy points of fine-tuned ESG classifiers, and beats the model's own answer in eleven of twelve comparisons.","lead":"The authors show that a simple linear probe reading a large language model's internal activations can judge whether financial text is about environmental, social, or governance topics almost as accurately as a classifier fine-tuned for that task. This suggests cheap, frozen-model monitoring can replace task-specific fine-tuning for concept measurement.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline 'within 0.6 points of fine-tuning' compares the best of nine probe configurations with the best of two fine-tuned models; no validation-based selection or multiplicity correction is reported, so the 0.6-point gap can be selection noise.","rationale":"The reader's weakest assumption focuses on the comparability of published fine-tuned accuracies with the new probe/RFM results across different evaluation protocols. That is a valid and important concern. My stress-test identifies a second, internal threat: even if the fine-tuned results were obtained on exactly the same splits, the headline gap is computed from the maximum of nine probe configurations, and no selection procedure or multiplicity adjustment is reported. This makes the 'within 0.6 points' claim unverifiable from the tables as presented. The concern is concrete, testable, and directly targets the central 'nearly as well as fine-tuning' assertion. It does not undermine the paper's stronger and better-controlled finding that activations carry concept content the output does not report, because that comparison is pairwise across all model-pillar combinations rather than best-of-N. The paper is transparent about several other limitations, including the test-split threshold sweep and wrapper sensitivity, but the best-of-nine selection is not flagged. I therefore keep the reader's CONDITIONAL verdict: the conclusion is plausible and partially supported, but the headline comparison needs a held-out configuration-selection check before it can be read as a statement about the method rather than about the luckiest configuration.","tokens_in":15787,"tokens_out":5371,"duration_ms":57369,"concrete_test":"For each pillar, fix a single held-out test split of 500 sentences before any configuration selection. Using only the training/validation portion, choose the linear-probe configuration (model, pooling strategy, wrapper) by validation accuracy; freeze it. Evaluate that one configuration on the held-out test set, and evaluate the fine-tuned baselines from Schimanski et al. (2024) on exactly the same held-out test set. If the pre-selected probe's accuracy is materially lower than the best-of-nine values in Tables 3–5, or the gap to fine-tuning widens beyond 0.6–2.1 points, the headline claim is selection-driven and should be rephrased as a property of the best observed configuration rather than of the method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that a frozen-model linear probe 'comes within 0.6, 1.0 and 2.1 accuracy points' of the best fine-tuned classifier (Abstract; Section 5) rests on a best-of-N comparison. In Tables 3–5, the reported 'best linear probing' entry is the maximum accuracy over 3 pooling strategies × 3 models, while the fine-tuned reference is the maximum over 2 architectures. The paper does not select the probe configuration on a validation split before seeing test results, and no multiple-comparison correction is applied. The reported ± values are per-configuration cross-validation standard deviations, not selection-adjusted uncertainties. On Environmental, the selected probe scores 0.951 ± 0.009 against EnvRoBERTa's 0.957 ± 0.010; the 0.6-point gap is smaller than one standard error and the expected maximum of nine positively correlated estimates is inflated upward. Consequently, the headline gap can arise even if every individual probe configuration is genuinely worse than fine-tuning. The like-for-like comparisons — probe vs RFM on the shared split and probe vs the model's own answer — do not suffer from this best-of-N selection and remain evidence for the internal-concept-content claim, but the 'stands in for fine-tuning' comparison is not established as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether the internal activations of frozen, out-of-the-box LLMs can measure concept content in text, using the ESG dataset of Schimanski et al. (2024) as a testbed. It compares two activation-based extractors—linear probing and RFM concept vectors—against an embedding baseline, surface baselines, the model's own yes/no answer, and the fine-tuned classifiers published with the dataset. The central reported result is that the best linear probe comes within 0.6, 1.0, and 2.1 accuracy points of the best fine-tuned model on the Environmental, Social, and Governance pillars, respectively, and beats the same model's own answer in eleven of twelve model-pillar combinations. The paper also reports that linear probing outperforms RFM in eleven of twelve like-for-like comparisons on a shared split, and it presents the RFM continuous score as an unvalidated graded measure.","tokens_in":15959,"tokens_out":5871,"duration_ms":53716,"significance":"If the headline claim were established, the paper would make a practically useful contribution: concept measurement without task-specific fine-tuning, at a few minutes of GPU time per pillar. The study has real strengths. It includes a like-for-like probe-versus-RFM comparison on an identical split, which the reader's preliminary report confirms as sound. It also tests the internal-versus-output premise directly by comparing activation-based measures with the model's own elicited answer, and it reports surface baselines (dictionary, topic model, TF-IDF) that put the activation-based results in context. The authors are admirably explicit about limitations: wrapper sensitivity, the unvalidated continuous score, the absence of Gemma from the probe tables, and the swept RFM threshold. The difficulty is that the strongest claim—that a frozen-model probe 'stands in for fine-tuning'—is not established by the current analysis because the comparison to fine-tuned classifiers is not a controlled one and the probe configuration is selected on the test outcome.","major_comments":[{"comment":"The headline 'within 0.6, 1.0 and 2.1 accuracy points' compares the maximum of nine probe configurations (3 pooling strategies × 3 models) with the maximum of two fine-tuned architectures, with no validation-based model selection or multiplicity correction. On Environmental, the selected probe scores 0.951±0.009 against EnvRoBERTa's 0.957±0.010, a gap smaller than one standard error, and the expected maximum of nine positively correlated estimates is inflated. Consequently, the paper does not establish that a linear probe on frozen activations is as accurate as a fine-tuned classifier; it establishes only that some probe configuration came close on the test split. The authors should either select the probe configuration on a validation split before seeing the test results, apply a multiple-comparison correction, or reframe the claim as an exploratory upper bound rather than a head-to-head result.","section":"Section 4.1 and Tables 3–5"},{"comment":"The comparison to the fine-tuned models of Schimanski et al. (2024) is not like-for-like. The probe is scored by 5-fold cross-validation over all 2,000 sentences, RFM is scored once on a 500-sentence held-out split, and the fine-tuned numbers are taken from the original paper without rerunning those models on the same splits. Differences in train/test composition or label distribution across protocols can easily produce gaps of the size reported. The authors should either rerun the fine-tuned models under the same evaluation protocol or clearly label the fine-tuned numbers as an external benchmark rather than a controlled comparison. This is load-bearing because the abstract and conclusion use the 'within 0.6 points' language as the paper's main justification for frozen-model monitoring.","section":"Section 4 and Tables 3–5"},{"comment":"The RFM decision threshold is swept on the split being reported, so the RFM accuracies and F1 scores in Tables 3–5 are the best attainable on that split rather than the values a frozen threshold would produce. The paper's own supplementary check reports that refitting the threshold on validation and applying it to the test split lowers accuracy by 0.011 on average and F1 by 0.025, with worst cases of 0.038 and 0.136. This optimistic reporting should be corrected: the tables should present the fixed-threshold values, and all claims about RFM's relative performance should be based on those values. The issue is acknowledged in the text, but it still affects every RFM number in the results and the 'seven of twelve' comparison with the model's own answer.","section":"Section 3.3.1 and Tables 3–5"}],"minor_comments":[{"comment":"The sentence 'the single exception is a tie (0.975 against 0.978 for Llama-3.1-8B on Environmental)' is internally inconsistent: 0.975 and 0.978 are not equal, so this is not a tie. The numbers should be reconciled or the wording corrected.","section":"Section 4.2"},{"comment":"The column headers contain intrusive spacing ('Y es (T rue)', 'No (F alse)', 'T otal'); this is a typesetting artifact that should be fixed.","section":"Table 2"},{"comment":"Several wrapper rows are visually similar and hard to distinguish. Assigning each wrapper a label (e.g., W1–W6) would make the 9.4-point sensitivity range easier to follow.","section":"Appendix A, Table 6"},{"comment":"Please clarify whether the precision, recall, F1, and AUC values for RFM are computed at the accuracy-optimizing threshold or at some other operating point, since the text describes the threshold sweep only in terms of accuracy.","section":"Section 3.3.1"},{"comment":"The paper says 'a few minutes of GPU time per pillar' but later quantifies the cost as under four minutes for RFM and about five minutes for the full probe. These statements are consistent but could be consolidated to avoid seeming imprecise.","section":"Section 5"},{"comment":"The paper does not state whether code and data will be publicly released. Given the reproducibility claims, an availability statement would be helpful.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope and addresses a genuinely useful question. The central difficulty is that the headline comparison to fine-tuned classifiers is not controlled, and the manuscript's own numbers show the gap is within one standard error even before accounting for best-of-N selection. I believe the authors can fix this by re-running the fine-tuned models under the same protocol or by substantially softening the claim. If they choose the latter, the paper would still contribute a clean probe-versus-RFM comparison and a novel test of the internal-knowledge premise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe like-for-like core of this paper is genuinely useful: linear probing beats RFM concept vectors in eleven of twelve model–pillar comparisons on the same split, and beats the same model's own yes/no answer in eleven of twelve. That's a real, controlled head-to-head that prior work didn't run. The headline claim — that the best probe comes within 0.6 percentage points of a fine-tuned classifier — is weaker than it reads. The 0.6 comes from picking the best of nine probe configurations (three models × three pooling strategies) and comparing to the best of two fine-tuned models from the original ESG paper, evaluated on different splits. No validation-based selection or multiplicity correction. The gap is smaller than the probe's own cross-validation standard error. So the 'stands in for fine-tuning' framing is optimistic; the 'activations know more than the output says' claim is well supported.\n\nCredit where due: the paper is transparent. It reports the RFM threshold sweep and its effect (freezing the threshold costs 0.011 accuracy on average, but up to 0.136 F1 in the worst cell), the 9.4-point wrapper sensitivity, the Gemma OOM, and explicitly says the continuous RFM score is unvalidated against graded labels. That's honest reporting.\n\nSoft spots in proportion: the threshold sweep and the best-of-N selection are fixable — select on validation, report all configurations or correct for selection, and ideally rerun the fine-tuned baselines on the same folds. The absence of code/data is a minor reproducibility hit. The wrapper sensitivity means absolute numbers are conditional on wrapper choice; the authors acknowledge it but don't solve it.\n\nThis deserves peer review. The qualitative result is solid and useful for anyone measuring concepts in text without fine-tuning. A serious referee should push for the selection fix and a softened headline. I'd cite the like-for-like comparison, not the 0.6-point number.","headline":"Useful head-to-head with an overstated headline: the probe-vs-RFM and probe-vs-output comparisons hold, but the 'within 0.6 points of fine-tuning' claim is best-of-N noise.","tokens_in":16593,"tokens_out":2609,"would_cite":true,"duration_ms":23321,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A linear probe on frozen LLM activations measures ESG concept content within 0.6 accuracy points of fine-tuned classifiers.","keywords":["LLM activations","linear probing","concept vectors","Recursive Feature Machine","ESG measurement","financial text","text-as-data","frozen models"],"falsifier":"Re-run the published fine-tuned ESG classifiers on the exact 5-fold cross-validation splits used for the linear probes; if the accuracy gap widens beyond 0.6 points or reverses on any pillar, the near-parity claim is not established.","tokens_in":15499,"feed_emoji":"🧠","tokens_out":7605,"duration_ms":55976,"temperature":0.7,"pith_summary":"The paper asks whether the internal activations of a frozen, out-of-the-box large language model can measure how much a text concerns a concept, without fine-tuning the model for that concept. On a human-annotated set of financial sentences labelled for the three ESG pillars, it finds that the answer is yes: a simple linear probe trained on layer activations classifies concept presence within 0.6 to 2.1 accuracy points of published fine-tuned domain classifiers. The same probe also beats the model's own yes/no answer in eleven of twelve comparisons, which supports the paper's central premise that activations carry concept content the response does not report. A more elaborate concept-vector method based on the Recursive Feature Machine trails the probe but produces a continuous score meant to reflect how strongly a concept is present, the validity of which remains untested because the dataset has only binary labels. The practical upshot is a cheap, fast measure of concept content that could substitute for fine-tuning when data or compute are scarce.","feed_headline":"Frozen LLM probe nears fine-tuned ESG classifiers","feed_subtitle":"Reading a frozen model's activations lands within 0.6 points of fine-tuned ESG classifiers, beating its yes/no answers.","key_machinery":"The machinery is linear probing of LLM activations, compared against RFM concept vectors. Each input sentence is passed through a frozen decoder-only model, and every layer's per-token hidden states are pooled (last-token, mean, or max) into one vector per layer; a ridge classifier is then fit on each layer's pooled activations, and a logistic regression meta-learner combines the per-layer decision scores via nested 5-fold cross-validation. The RFM alternative computes an Average Gradient Outer Product matrix from labelled activations and takes its top eigenvectors as concept directions, scoring new sentences by cosine similarity and thresholding the score. Token pooling and a task-framing wrapper are varied, and the best wrapper is selected on a validation split. The comparison against the embedding baseline, surface baselines, the model's own answer, and published fine-tuned classifiers carries the argument.","core_discovery":"The central claim is that a linear probe on the frozen activations of an off-the-shelf LLM measures ESG concept presence nearly as well as a classifier fine-tuned for the task. On the Environmental pillar the best probe reaches 0.951 accuracy against 0.957 for the best fine-tuned model; on Social it reaches 0.924 against 0.934; on Governance 0.876 against 0.897. The probe outscores the same model's own answer to the same binary question in eleven of twelve model-pillar combinations, by 0.043 AUC on average, so the internal signal is not merely echoing the output. Against the other activation-based extractor, the probe wins in eleven of twelve like-for-like comparisons even when both are trained on identical splits. The paper therefore proposes the frozen-model linear probe as a serious standing alternative to task-specific fine-tuning for measuring concept content.","pith_inferences":["Editorial inference: the same probe-on-frozen-activations recipe should transfer to other regulated-text domains, such as risk disclosure, corporate culture, or contract clauses, where labelled sets are small and fine-tuning is expensive; the ESG result is evidence of mechanism, not a domain-bound accident.","Editorial inference: because the probe outperforms the model's verbal yes/no answer, activation-based reading could serve as a monitoring channel for machine-generated corporate text, where surface wording and underlying concept content may drift apart.","Editorial inference: the consistent superiority of the first RFM eigenvector over the mean of five suggests the AGOP concept vectors are capturing one dominant separation; inspecting what the second through fifth eigenvectors separate would clarify whether averaging adds signal or noise.","Editorial inference: a direct test of the continuous-score claim would re-annotate a sample of ESG sentences on an ordinal scale, such as weak versus strong presence, and rank-correlate the RFM cosine scores against those grades; the paper already notes such validation awaits graded labels."],"forward_implications":["Firms and researchers can measure concept presence in text with a few minutes of GPU time per concept and no task-specific fine-tuning, which lowers the cost of large-scale ESG and financial text monitoring.","Because the probe beats the model's own answer, downstream applications that rely on LLM outputs alone may miss concept content that is accessible from internal states; reading activations can recover that content.","The RFM continuous score, if validated on graded labels, would extend measurement from classification (present or not) to degree of concept presence, giving text-as-data researchers a graded measure.","The 9.4-point accuracy swings from changing the prompt wrapper imply that any practitioner using these methods must treat wrapper design as a first-order hyperparameter, not a fixed detail."],"supporting_citations":[{"why":"Supplies the annotated ESG dataset and the published fine-tuned classifier accuracies that serve as the reference point.","marker":"Schimanski et al. (2024)"},{"why":"Introduces RFM-based concept vectors and the universal steering and monitoring approach this paper tests.","marker":"Beaglehole, Radhakrishnan, et al. (2026)"},{"why":"Defines the Recursive Feature Machine algorithm and the AGOP matrix from which concept vectors are extracted.","marker":"Beaglehole, Holzmüller, et al. (2026)"},{"why":"Introduces linear classifier probes for reading intermediate layers, the method this paper uses.","marker":"Alain & Bengio, 2017"},{"why":"Surveys probing classifiers and grounds the interpretation of probe predictions as concept presence.","marker":"Belinkov, 2022"},{"why":"States the linear representation hypothesis that concepts are linearly accessible directions in activation space.","marker":"Marks & Tegmark, 2024"},{"why":"Formalises the linear representation hypothesis for LLMs, underpinning the concept-direction premise.","marker":"Park et al., 2024"},{"why":"Shows linear probes can read graded quantities like place and time from frozen activations, motivating graded concept measurement.","marker":"Gurnee & Tegmark, 2024"},{"why":"Supplies the pooling strategies and embedding baseline approach used for token pooling.","marker":"Reimers & Gurevych, 2019"},{"why":"Projects frozen LLM hidden states onto learned concept directions for continuous text measures, the closest prior application.","marker":"Yang et al. (2024)"}],"fun_headline_variants":["Frozen LLM probe nears fine-tuned ESG accuracy","Activation probe outscores LLM's own ESG answers","Linear probe on frozen LLM rivals fine-tuning","No fine-tuning needed: LLM activations read ESG concepts","Internal LLM signal beats its surface ESG judgments"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the published fine-tuned classifier accuracies are directly comparable to the probe and concept-vector numbers even though the fine-tuned models were not re-run on the same data splits; if the evaluation protocols differ in train/test composition or label distribution, the claimed 0.6-to-2.1-point margins do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Frozen LLM probe nears fine-tuned ESG accuracy","Activation probe outscores LLM's own ESG answers","Linear probe on frozen LLM rivals fine-tuning","No fine-tuning needed: LLM activations read ESG concepts","Internal LLM signal beats its surface ESG judgments"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1662,"prompt_tokens":1008,"completion_tokens":654,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":624,"completion_tokens_details":{"reasoning_tokens":575}},"tokens_in":624,"tokens_out":654,"duration_ms":6536,"temperature":1.0,"reasoning_tokens":575,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T12:47:45.589790+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the published fine-tuned ESG classifiers on the exact 5-fold cross-validation splits used for the linear probes; if the accuracy gap widens beyond 0.6 points or reverses on any pillar, the near-parity claim is not established.","supporting_citations":[{"cited_title":", Reding, A","cited_arxiv_id":null,"evidence_quote":"Supplies the annotated ESG dataset and the published fine-tuned classifier accuracies that serve as the reference point."},{"cited_title":", Holzm \\\"u ller, D","cited_arxiv_id":null,"evidence_quote":"Introduces RFM-based concept vectors and the universal steering and monitoring approach this paper tests."},{"cited_title":", Holzm \\\"u ller, D","cited_arxiv_id":null,"evidence_quote":"Defines the Recursive Feature Machine algorithm and the AGOP matrix from which concept vectors are extracted."},{"cited_title":"\\ Bengio, Y","cited_arxiv_id":null,"evidence_quote":"Introduces linear classifier probes for reading intermediate layers, the method this paper uses."},{"cited_title":"\\ Tegmark, M","cited_arxiv_id":null,"evidence_quote":"States the linear representation hypothesis that concepts are linearly accessible directions in activation space."},{"cited_title":", Choe, Y J","cited_arxiv_id":null,"evidence_quote":"Formalises the linear representation hypothesis for LLMs, underpinning the concept-direction premise."}],"review_version":1}