{"id":"2e567f79-bf75-4f5f-9532-f50f35a05cef","arxiv_id":"2505.16491","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LLaMA models encode binary sentiment most strongly in middle layers and emotions in early layers, and truncating the model at the best layer with a probe head yields efficient sentiment classifiers.","lead":"This paper probes the internal layers of LLaMA language models to find where positive and negative sentiment and six emotions are encoded, then uses that location to build a smaller, faster sentiment classifier. It reports that sentiment peaks in middle layers for polarity, emotions in early layers, and that a truncated model with a lightweight head beats prompting and approaches DeBERTa and RoBERTa accuracy with fewer parameters.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SENTRILLAMA's accuracy and the layer-localization findings are selected on the same test set used for comparison; the reported gains and 57% memory reduction are therefore optimistic until confirmed on a validation split.","rationale":"The reader's weakest-assumption analysis already identifies test-set-based selection as the critical flaw. I read the paper in good faith: the probing methodology is clearly described, the code is released, and the qualitative finding that the last token is not always best is plausible. The load-bearing numerical claims, however, depend on choosing the best of many configurations on the same test set used for the final comparison. This is not an internal inconsistency in the probing equations, but a correctness risk in the inference from Table 3 to Table 4. The recommended validation-split check is small and would either confirm or refute the claims; hence the verdict should remain CONDITIONAL as the reader stated. I see no additional independent objection strong enough to move the verdict.","tokens_in":20229,"tokens_out":3857,"duration_ms":34759,"concrete_test":"Reproduce the SENTRILLAMA pipeline with an untouched test split: use only training data (or a separate validation split) to select the layer, pooling, and probe for each model/dataset, then evaluate once on the held-out test set. If the validation-selected test accuracy is materially below Table 4 (e.g., more than 1-2 points on SST-2/IMDB), or the selected layer shifts, the SENTRILLAMA and localization claims are unsupported. Add repeated-seed or bootstrap confidence intervals to the selected configuration to quantify selection bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central applied claim (SENTRILLAMA approaches DeBERTa/RoBERTa while cutting parameters) is built from Table 3, which reports the top-3 layer/pooling/classifier configurations ranked by test accuracy. Section 5 then defines the 'most representative layer' as that same test-selected row and reports it again in Table 4. The search space is roughly 15-32 layers times 6 pooling methods times 12 classifiers, so choosing the maximum over this grid on the held-out test set inflates expected accuracy and invalidates the comparison against DeBERTa/RoBERTa and prompting in Table 4. The localization claims ('mid-layers for polarity, early layers for emotion') are also post-hoc selections, not predictions; without a validation split they do not establish where sentiment is 'most represented.' No error bars or repeated-seed statistics are reported. The paper itself (Discussion) notes the layer-selection mechanism is under-explored but does not flag the test-set selection issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper probes Llama-3 (1B, 3B, 8B, instruction-tuned) hidden states for binary sentiment and six-way emotion classification. It compares six pooling methods and twelve probe classifiers across layers, reports where sentiment/emotion information is 'most represented,' proposes SENTRILLAMA, a model truncated to the most representative layer plus a lightweight classifier, and argues that this approach outperforms prompting, approaches DeBERTa/RoBERTa on some datasets, and reduces memory usage by 57% on average. The central scientific claims are that binary sentiment is best detected in mid layers, emotions in early layers, that last-token pooling is not consistently best, and that layer-specific probing is a practical alternative to prompting.","tokens_in":20464,"tokens_out":3479,"duration_ms":29561,"significance":"If the claims are correct, the paper offers a practical, parameter-efficient recipe for sentiment classification and a useful map of where affective information resides in decoder-only LMs. The study has real strengths: a broad experimental grid across three model scales and four datasets, six pooling methods, twelve classifiers, shared code (stated), and deterministic seed configuration. The qualitative picture is suggestive and largely consistent across models and datasets. However, the headline quantitative claims rest on a protocol that selects the best layer/pooling/probe on the test set and then reports accuracy on that same test set, so the magnitudes of the claimed gains and the SENTRILLAMA comparisons are not established as stated. The contribution is potentially valuable but needs a corrected evaluation protocol before its central quantitative conclusions can be accepted.","major_comments":[{"comment":"The central quantitative results are selected on the test set. Section 4 states that probe accuracy is measured on the test set, Table 3 is built by ranking configurations by test accuracy, and Section 5 defines the 'most representative layer' L_i as that same test-selected row, which Table 4 then reports as SENTRILLAMA's accuracy on the same test set. With a search space of roughly 15–32 layers × 6 pooling methods × 12 classifiers, taking the maximum on the held-out test set inflates the expected accuracy and invalidates the comparison against prompting, DeBERTa, and RoBERTa in Table 4. Please redo the selection with a proper validation split (or nested cross-validation), then report test accuracy only for the configuration selected on validation, with confidence intervals or repeated-seed statistics.","section":"§4, §5, Tables 3–4"},{"comment":"The abstract's claim of 'detection accuracy increasing up to 14% over prompting techniques' is not traceable to Table 4. The reported SENTRILLAMA-versus-prompting gaps range from a few points (e.g., 0.9594 vs. 0.9154 on SST-2 for the 3B model) to more than 45 points (e.g., 0.8015 vs. 0.3475 on Emotion for the 1B model). No row yields a 14% gap. Either identify the exact comparison being summarized or correct/remove the claim.","section":"Abstract, §5, Table 4"},{"comment":"The localization claims ('mid-layers for binary sentiment, early layers for emotions') are post-hoc selections rather than independent predictions. Because the best layer is chosen by maximizing test accuracy and the paper reports a single split with no error bars, the qualitative layer trends are suggestive but not statistically supported. Please add a validation-based selection procedure and report per-layer confidence intervals (bootstrap or repeated seeds) so the localization pattern can be distinguished from selection noise.","section":"§4, §6, Conclusion, Figures 1 and 6"},{"comment":"The comparison against DeBERTaV3-large and RoBERTa-large is not fully specified. The paper does not state whether these baselines are fine-tuned on the same reduced training splits used for SENTRILLAMA or on the original full IMDB and Emotion datasets. If the baselines use the full datasets while SENTRILLAMA is selected on the test set, the comparison is doubly favorable to SENTRILLAMA. Please specify the baseline training protocol, and ideally fine-tune the baselines on the same reduced splits used for the proposed method.","section":"§5, Table 4, Appendix A"}],"minor_comments":[{"comment":"The phrase 'we adopt the residuals stream' contains a typo; it should read 'the residual stream.'","section":"§3, Choice of Representation"},{"comment":"The heading 'Compuattional Efficiency of SENTRILLAMA' contains a typo; it should be 'Computational Efficiency.'","section":"Appendix G"},{"comment":"The caption for subfigure (g) reads '1B-Instruct on Rotten Rotten'; the dataset name should be 'Rotten Tomatoes.'","section":"Appendix F, Figure 7"},{"comment":"The code availability footnote states 'Sentiment Probing Toolkit' but does not include a URL. Please provide a working link or repository identifier.","section":"§3, footnote 1"}],"recommendation":"major_revision","confidential_remarks":"The test-set selection issue is the main obstacle. It is a standard protocol flaw, not a philosophical disagreement, and the authors' stated code release and fixed seed configuration make the repair feasible. If a rerun with a validation split substantially changes Tables 3 and 4, the localization and efficiency claims should be scaled back accordingly. The qualitative layer trends in the figures suggest the scientific question is worth pursuing, but the current manuscript's headline numbers cannot be taken at face value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the qualitative findings are probably solid, but the headline numbers are inflated by selecting the layer, pooling, and classifier on the test set; still worth a serious look.\n\nThe genuinely new part is the systematic layer-wise map of sentiment and emotion in Llama-3 (1B, 3B, 8B) using six pooling methods and twelve classifiers, plus SENTRILLAMA, a truncated model that keeps only the layers up to the most informative one and attaches a lightweight head. That's a useful practical recipe, and the observation that the last token is not always the best sentence representation is a good corrective to common practice.\n\nThe paper does several things well. The code is released. The qualitative curves are consistent across model sizes and datasets: binary polarity peaks in mid-layers, emotions peak in early layers. Those patterns are visible in every plot, so they're unlikely to be artifacts. The efficiency numbers (memory, throughput) are concrete and the comparison against the full Llama is fair.\n\nThe soft spot is central and real: the best layer, pooling, and classifier are selected by maximizing accuracy on the test set, and then that same test-set accuracy is reported as SENTRILLAMA's performance and compared against DeBERTa, RoBERTa, and prompting. That's a fitted selection, not an independent result, and it makes the reported gains optimistic. There's no validation split and no error bars. The abstract's 'up to 14%' improvement over prompting doesn't map cleanly to any table. Also, the DeBERTa/RoBERTa baselines are underspecified: were they fine-tuned on the same reduced training sets? That matters for a fair comparison.\n\nThese flaws are fixable. A validation split for selection, repeated seeds, and a cleaner baseline description would likely leave the qualitative findings intact. The mid-layer/early-layer story is robust enough that I'd expect it to survive.\n\nThis paper is for people working on LLM interpretability and on cheap adaptation of open models for classification. It's not a breakthrough, but it's a legitimate contribution with a clear practical angle. I'd engage with it, and if I were reviewing it, I'd ask for the validation-split re-run before trusting the numbers.","headline":"The qualitative layer map of sentiment in LLaMA is probably right, but the headline numbers are inflated by test-set selection; still worth a serious look.","tokens_in":20976,"tokens_out":2994,"would_cite":false,"duration_ms":24232,"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":"The paper claims that sentiment in LLaMA models is most detectable in mid-layer residual streams, emotions in early layers, and that a LLaMA truncated to its most representative layer plus a lightweight probe beats prompting.","keywords":["sentiment analysis","emotion classification","probing classifiers","LLaMA","layer-wise analysis","pooling strategies","model truncation","efficient inference"],"falsifier":"A concrete control would be a label-shuffle test: train the same probes on the same residual-stream features with randomly permuted sentiment labels across all layers. If the mid-layer accuracy peak for binary polarity survives at high levels, then the reported localization is an artifact of the probe, pooling, and layer-selection procedure rather than a genuine concentration of sentiment information.","tokens_in":20068,"feed_emoji":"🔬","tokens_out":5010,"duration_ms":37214,"temperature":0.7,"pith_summary":"The paper asks where sentiment and emotion information lives inside LLaMA models and whether that location can be put to practical use. By training probe classifiers on residual-stream representations at each layer, it argues that binary positive/negative polarity is most detectable in middle layers, while fine-grained emotions are most detectable in early layers. It also claims that the common practice of using the last token as a sentence summary is suboptimal, with a concatenation of mean, max, and min pooling often doing better. Building on the layer map, the paper introduces SENTRILLAMA, a LLaMA truncated to the most representative layer and fitted with a lightweight classification head, which it reports matches or beats prompting-based LLaMA and approaches fine-tuned DeBERTa and RoBERTa accuracy while reducing memory use. A sympathetic reader would care because the result offers both a map of affective information in decoder-only transformers and a recipe for cheaper sentiment classifiers.","feed_headline":"Probing shows sentiment lives in LLaMA's mid-layers","feed_subtitle":"Trimming LLaMA to its best layer beats prompting and nears DeBERTa accuracy with less memory.","key_machinery":"The central object is the probe: a classifier (typically a linear SVM or logistic regression) trained on the residual-stream representation $x^{(l+1)}$ at a given layer $l$, pooled over tokens by one of six methods (last-token, mean, max, min, concatenation of mean-max-min, and attention-weighted mean). A forward pass through the LLaMA model produces the representation, the probe is trained on labeled sentiment data, and the layer with the highest test accuracy is taken to be where sentiment is most represented. SENTRILLAMA then keeps only the embedding layer and the transformer layers up to that best layer, replaces the language-model head with a lightweight classifier, and thereby defines a smaller model that inherits the sentiment-relevant representations.","core_discovery":"The central claim is that sentiment and emotion concepts are encoded in LLaMA's hidden layers in a localizable way: linear classifiers reach roughly 90 percent accuracy on binary polarity when reading the residual stream at middle layers, and around 70 percent for six-way emotion classification when reading early layers. The paper further claims that the last token is not consistently the best pooling choice, and that a truncated model retaining only layers up to the most representative one plus a linear SVM matches or exceeds zero-shot, few-shot, and chain-of-thought prompting while approaching DeBERTa-large and RoBERTa-large performance with an average parameter reduction of about 57 percent. The reported best layer moves deeper as the model grows, from around layer 8 of 16 in the 1B model to around layer 20 of 32 in the 8B model on SST-2.","pith_inferences":["If the localization holds, the same truncation recipe could extend to other classification tasks with labeled data, such as sarcasm or intent detection, though the paper does not test this.","If sentiment is indeed most readable in a single mid-layer, activation-steering and representation-engineering methods that target one layer could be aimed at that identified layer for more effective interventions.","A direct test of the localization claim would be a label-shuffle control: if the mid-layer accuracy peak survives when probe labels are randomly permuted, the result is far more likely to reflect genuine representation structure rather than dataset artifacts."],"forward_implications":["Layer-specific probing can serve as an alternative to prompting for sentiment classification, with reported accuracy gains of up to 14 percent.","Truncating LLaMA at the most representative layer cuts memory requirements by roughly 57 percent on average while retaining most of the accuracy, making large-model sentiment classifiers cheaper to deploy.","Pooling choice matters: concatenating mean, max, and min representations often outperforms the last token, which suggests that studies relying solely on last-token representations may be leaving signal unused.","Larger models encode sentiment more cleanly, with the best layer appearing deeper, so scale and layer choice interact when designing downstream probes."],"supporting_citations":[{"why":"Defines probe classifiers, the core method used to read out sentiment from hidden layers.","marker":"(Alain and Bengio, 2017)"},{"why":"Provides the concept detection framework that the sentiment probing setup adapts.","marker":"(von Rütte et al., 2024)"},{"why":"Prior example of probing latent knowledge from language models, using last-token representations as a baseline this paper challenges.","marker":"(Burns et al., 2023)"},{"why":"Establishes the line of probing internal states for honesty, extended here to sentiment.","marker":"(Azaria and Mitchell, 2023)"},{"why":"Prior probing of factual knowledge using last-token representations, another baseline the paper contrasts with its pooling alternatives.","marker":"(Li et al., 2023)"},{"why":"Source of the prompt templates used for zero-shot, few-shot, and chain-of-thought baselines.","marker":"(Deng et al., 2023)"},{"why":"Introduces the Llama architecture that is probed in this work.","marker":"(Touvron et al., 2023)"},{"why":"Introduces the Llama-3 architecture with Grouped Query Attention and RMSNorm used here.","marker":"(Dubey et al., 2024)"},{"why":"DeBERTaV3-large, the fine-tuned baseline that SENTRILLAMA approaches.","marker":"(He et al., 2023)"},{"why":"RoBERTa-large, another fine-tuned baseline that SENTRILLAMA matches or approaches.","marker":"(Liu et al., 2019)"}],"fun_headline_variants":["Sentiment hides in LLaMA mid-layers, probes reveal","Probing LLaMA shows sentiment is mid-layer, memory drops 57%","LLaMA's feelings are in its middle layers, probes show","Mid-layer probes beat prompting for LLaMA sentiment","LLaMA sentiment is best read from mid-layers, not last token"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that probe accuracy at the most representative layer reflects where sentiment is actually encoded in the model, rather than dataset regularities or the probe-selection procedure itself; the paper selects the best layer, pooling, and classifier on the test set, so the localization claim is not an independent prediction.","fun_headline_variants_meta":{"raw":{"variants":["Sentiment hides in LLaMA mid-layers, probes reveal","Probing LLaMA shows sentiment is mid-layer, memory drops 57%","LLaMA's feelings are in its middle layers, probes show","Mid-layer probes beat prompting for LLaMA sentiment","LLaMA sentiment is best read from mid-layers, not last token"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001111,"raw_usage":{"total_tokens":4611,"prompt_tokens":911,"completion_tokens":3700,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":3607}},"tokens_in":527,"tokens_out":3700,"duration_ms":19075,"temperature":1.0,"reasoning_tokens":3607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:59:40.831014+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete control would be a label-shuffle test: train the same probes on the same residual-stream features with randomly permuted sentiment labels across all layers. If the mid-layer accuracy peak for binary polarity survives at high levels, then the reported localization is an artifact of the probe, pooling, and layer-selection procedure rather than a genuine concentration of sentiment information.","supporting_citations":[],"review_version":1}