{"id":"b8b797c3-4d31-44df-ad29-32606144f3a5","arxiv_id":"2508.17297","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A sparse autoencoder with synthetic user probes identifies popularity-encoding neurons in a recommender model, and steering those neurons improves exposure fairness with limited accuracy loss.","lead":"This paper presents PopSteer, a method that uses sparse autoencoders to locate and adjust the internal signals in a recommendation model that favor popular items. The goal is to make recommendations fairer to long-tail items while preserving accuracy and adding a degree of model interpretability.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Synthetic user generation in Section 3.1 cannot isolate popularity from item-content and sequence-coherence confounds, undermining the causal claim that steered neurons encode popularity rather than correlated features; the effectiveness result survives, but interpretability remains unvalidated.","rationale":"The reader's weakest assumption—that random replacement of items may introduce content and temporal confounds—is exactly the most load-bearing gap in the paper. The synthetic profiles are the sole basis for neuron selection, and the neuron-level interpretability claim rests entirely on the assumption that the only systematic difference between R_pop and R_unpop is item popularity. That assumption is insecure because popularity is strongly correlated with genre, recency, and sequence coherence in both ML-1M and Last.fm. If the confound is real, the method may still improve fairness by suppressing 'mainstream' content, but the paper's central framing as an interpretable, neuron-level explanation of popularity bias would be misleading, and the method's behavior on other datasets could vary unpredictably with those correlations. I do not find this fatal to the empirical effectiveness claim: the random-noise ablation provides some evidence that targeted steering is genuinely doing something, and the results are consistent across two datasets. But it is a condition for accepting the interpretability contribution, which is the paper's main novelty. The proposed real-user group overlap test and genre-blocked rerun would directly settle the issue. Because the reader already flagged this assumption and asked for follow-up, and because no additional zero-confidence concern emerged, the CONDITIONAL verdict stands unchanged.","tokens_in":8699,"tokens_out":6239,"duration_ms":64699,"concrete_test":"Recompute Cohen's d using real-user groups instead of synthetic profiles: split each dataset's users by the popularity share of their interacted items (e.g., top tercile vs bottom tercile), feed real sequences through the same pretrained SASRec+SAE, compute d_j for all neurons, and measure the Jaccard overlap between the top-N neurons selected this way and those selected by the synthetic method of Section 3.1. If the overlap is below 0.5 or the sign of d_j is inconsistent for the overlapping neurons, the synthetic replacement introduces confounds and the 'popularity neuron' interpretation is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of PopSteer depends on Section 3.1, where each synthetic profile replaces every item in a real user's history with a randomly drawn item from I_pop or I_unpop. This preserves sequence length but destroys the original item content, genre, recency, and temporal ordering. Because SASRec is a sequential model, an all-popular sequence is not just 'popularity-labeled' but also more homogeneous, more mainstream in genre/artist, and temporally atypical. The SAE activations on these out-of-distribution sequences can differ from activations on all-unpopular sequences for reasons unrelated to popularity per se. Cohen's d then selects neurons that may encode 'mainstream content', 'genre typicality', or 'sequence predictability' rather than a popularity feature. The interpretability analysis in Section 5.3.1 ('deactivating popularity-positive neurons decreases Gini') reuses the same synthetic-selected neurons, so it does not independently validate the popularity interpretation. The random-noise ablation (Section 5.3.2) only establishes that targeted selection beats random perturbation at equal nDCG budget; it does not establish that the target is popularity. Since the paper's claimed contribution is an *interpretable* remedy, this confound is the weakest load-bearing point: the fairness gains may be real but the explanation of why they occur, and the generalization of the method to datasets with different popularity-content correlations, is left unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PopSteer, a post-hoc interpretability and debiasing method for sequential recommender systems. A sparse autoencoder (SAE) is trained to reconstruct the user embedding produced by a pretrained SASRec model. Synthetic user histories are constructed by replacing every item in a real user's history with a random item from either the popular or the unpopular item set; Cohen's d between SAE activations on the two synthetic sets is used to identify 'popularity neurons.' PopSteer then adjusts the activations of the highest-|d| neurons, suppressing positive-d neurons and boosting negative-d neurons, and the modified user embedding is used to generate recommendations. Experiments on ML-1M and Last.fm compare PopSteer with IPR, FA*IR, PCT, P-MMF, and a random reranker, and an ablation compares targeted steering with random Gaussian activation noise. The paper reports improved long-tail coverage and Gini index with small nDCG@10 loss, and a neuron-deactivation study that links the selected neurons to changes in the Gini index.","tokens_in":8963,"tokens_out":5843,"duration_ms":59779,"significance":"If the interpretability claim is validated, PopSteer would be a practically useful post-hoc tool: it is model-agnostic at the embedding level, offers a fine-grained fairness--accuracy trade-off, and ships with code. The paper also deserves credit for comparing against several established baselines and for including a random-noise ablation, which is a good control for asking whether the steering mechanism itself matters. The central effectiveness result appears plausible. However, the paper's distinctive contribution is the claim that the steered neurons encode popularity specifically, and that claim currently rests on a synthetic-probe construction with a potential content confound. The 'significantly' and 'consistently' statements in the results are also stronger than the reported point estimates support. With additional validation, the method would be a solid contribution to interpretable fairness in recommendation.","major_comments":[{"comment":"The synthetic-probe construction does not isolate popularity from item content and sequence coherence. Replacing every item in a user's history with a randomly selected item from I_pop or I_unpop preserves only the profile length; it destroys genre, artist, recency, and the original temporal ordering, all of which are properties that a sequential model like SASRec can exploit. An all-popular synthetic sequence and an all-unpopular synthetic sequence therefore differ in many dimensions besides popularity. The Cohen's d selection in Eq. (2) may consequently identify neurons that encode 'mainstream content,' 'genre typicality,' or 'sequence predictability' rather than popularity per se. The interpretability analysis in Section 5.3.1 reuses exactly these synthetic-selected neurons, so it does not independently validate the popularity interpretation, and the random-noise ablation in Section 5.3.2 only establishes that targeted selection beats unstructured perturbation. I recommend adding a validation experiment that controls for content, for example by matching genres/categories and recency distributions between the popular and unpopular probe sets, or by testing whether the selected neurons predict popularity on held-out natural user histories.","section":"Section 3.1 and 3.2"},{"comment":"The claim that 'PopSteer consistently yields superior fairness outcomes relative to random Gaussian noise perturbation' is contradicted by the reported numbers. In ML-1M with N=1024, random noise achieves a lower Gini index (0.7387 vs. 0.7470) and higher long-tail coverage (0.5333 vs. 0.5065) than PopSteer; with N=2048, random noise also has higher long-tail coverage (0.5440 vs. 0.5346). Since lower Gini and higher coverage are better, the noise baseline is at least as fair in those configurations. The comparison is also not accuracy-matched: at N=1024, PopSteer has nDCG@10 of 0.1169 while noise has 0.1123. The ablation should be re-run with matched accuracy (or a proper Pareto-front comparison), and the 'consistently' claim should be tempered.","section":"Section 5.3.2, Table 2"},{"comment":"The steering equation uses sigma_j without defining it. Eq. (2) defines sigma_j,pop and sigma_j,unpop separately, but Eq. (4) refers to a single sigma_j. Is this the pooled standard deviation, one of the two group standard deviations, or another quantity? The magnitude of the activation adjustment depends directly on this choice, so the method as written is not fully specified and the implementation is not reproducible without this detail.","section":"Section 4, Eq. (4)"},{"comment":"The text says PopSteer 'significantly improves fairness' and 'consistently outperforms all baselines,' but no variance estimates, confidence intervals, or significance tests are reported anywhere in the paper. The reported points are single values after per-dataset hyperparameter tuning, which does not support 'significantly' in a statistical sense. At minimum, multiple seeds with error bars and a paired significance test should be reported; otherwise the wording should be weakened to describe the observed improvements in these experiments.","section":"Section 5.2, Figures 1 and 2"}],"minor_comments":[{"comment":"The symbol N is used both for the total number of SAE hidden neurons and for the number of steered neurons selected by PopSteer, which is confusing. Please use distinct notation, for example N_hidden and N_steer.","section":"Section 4 and Section 5.2"},{"comment":"The phrase 'both compromising roughly 20% of total interactions' should be 'comprising'; more importantly, the definition of the 20% threshold is ambiguous. Please specify whether I_pop and I_unpop are defined by cumulative interaction share, by rank cutoff, or by another rule.","section":"Section 3.1"},{"comment":"There is a typo: 'timestamp informtion' should be 'timestamp information.'","section":"Section 5.1"},{"comment":"The text writes 'ReLu' where it should be 'ReLU.'","section":"Section 2"},{"comment":"The SAE hyperparameters s and K are only given as search ranges; the final chosen values for each dataset are not reported. Please report the settings used in the main experiments.","section":"Section 5.2"},{"comment":"The number K' of deactivated neurons in the interpretability analysis is never defined, and Figure 3 does not show numerical values. Please specify the range of K' and, if possible, include the axis values or a small table.","section":"Section 5.3.1"},{"comment":"The auxiliary loss L_aux is referenced but not defined in the text; if space is a constraint, a one-sentence description or a precise citation to the formula in [8] would improve reproducibility.","section":"Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the interpretability claim, which is central to the paper's title and framing, rests on a synthetic-probe construction that has not been validated against content confounds. The effectiveness result is plausible, but the ablation claim is overstated in at least one configuration of Table 2. I believe the paper can be brought to publishable form with targeted additional experiments: a content-matched probe validation, accuracy-matched noise comparisons, and a clearer definition of Eq. (4). The lack of error bars is also worth addressing before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper transfers SAE feature steering from LLMs to sequential recommenders: it trains a sparse autoencoder on SASRec's user embeddings, builds synthetic popular/unpopular profiles to pick neurons via Cohen's d, and then adjusts activations to reduce popularity bias. On ML-1M and Last.fm the method improves long-tail coverage and Gini while keeping nDCG@10 close to the unsteered model, and the random-noise ablation makes a solid case that the targeted selection is what drives the fairness gains. The effectiveness result looks real.\n\nWhat's genuinely new is the packaging: SAE-based interpretability plus a post-hoc steering knob for popularity bias in recommenders. Each ingredient is established, but the combination is new and practically useful—no retraining, fine-grained control over the fairness-accuracy trade-off. The paper is clearly written and easy to follow.\n\nThe soft spot is the interpretability claim, which is under-validated. The synthetic profiles in Section 3.1 replace every item in a user's history with a random popular or unpopular item. That preserves sequence length but destroys item content, genre, and temporal coherence. SASRec is sequential, so an all-popular history is not just 'popular'—it's also more homogeneous and temporally atypical. Cohen's d then picks neurons that may respond to these correlated features rather than to a general popularity signal. The manipulation study in 5.3.1 reuses the same synthetic-selected neurons, so it doesn't independently confirm the popularity interpretation. The random-noise ablation only shows targeted steering beats random perturbation, not that the target is truly popularity. So the method works, but the explanation of why it works is not nailed down. If popularity correlates differently with content in another domain, the selected neurons may not transfer. That's a limitation, not a fatal flaw.\n\nMinor caveats: no error bars or significance tests, and alpha and N are tuned per dataset, so the reported points are best-case. The figures would be more convincing with error bars or at least multiple runs.\n\nBottom line: this is a useful contribution to the recommender fairness subfield. It deserves a serious referee. If I were reviewing, I'd ask the authors to validate the interpretability claim—e.g., control for genre/temporal features in the synthetic probes or show that selected neurons transfer across datasets. But the core effectiveness result probably stands.\n\nRecommendation: send to peer review, with a request for that validation.","headline":"PopSteer is a clever, effective post-hoc steering method for popularity bias, but the neuron-level interpretability claim rests on synthetic probes that don't isolate popularity from content correlations.","tokens_in":9491,"tokens_out":3345,"would_cite":true,"duration_ms":37024,"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":"PopSteer shows that popularity bias in a sequential recommender can be attributed to specific sparse-autoencoder neurons and corrected by steering their activations.","keywords":["popularity bias","recommender systems","sparse autoencoder","neuron steering","interpretability","fairness","sequential recommendation"],"falsifier":"Generate synthetic profiles in which popular and unpopular items are matched on genre, category, and release-era distributions before computing the per-neuron effect sizes; if the neurons identified as popularity-positive stop predicting the Gini-index improvement from steering, the attribution is driven by those confounds and not by popularity alone.","tokens_in":8456,"feed_emoji":"⚖️","tokens_out":7244,"duration_ms":68761,"temperature":0.7,"pith_summary":"PopSteer is a post-hoc method for opening up a deep recommender's popularity bias: it trains a sparse autoencoder (SAE) on the embeddings of a pretrained sequential recommender, then uses synthetic user profiles built from popular and unpopular items to find which SAE neurons respond to popularity. The paper claims that suppressing neurons that fire for popular items and amplifying neurons that fire for unpopular items improves exposure fairness, measured by long-tail coverage and the Gini index, while keeping nDCG@10 close to the original model. The point matters because existing debiasing methods reweight scores or rerank lists without explaining what inside the model drives the bias. PopSteer offers both a diagnosis, which neurons encode popularity, and a fine-grained control knob for the fairness-accuracy trade-off.","feed_headline":"Neuron steering fixes popularity bias with minimal accuracy loss","feed_subtitle":"Sparse autoencoder finds the neurons behind popularity bias and steers them to fairer lists with minimal accuracy loss.","key_machinery":"The central object is a sparse autoencoder attached to the final layer of a pretrained sequential recommender. It reconstructs each user embedding $\\mathbf{x}$ as $\\hat{x} = W_{\\mathrm{dec}} a + b_{\\mathrm{pre}}$ with a hidden activation vector $a$ that keeps only its top-$K$ largest entries, so each active neuron is forced to specialize. The machinery has three moving parts: the synthetic popular/unpopular user profiles used to probe the SAE; a per-neuron standardized effect size $d_j$ computed from the activation difference between the two profile types; and the steering rule $a'_j = a_j \\pm w_j \\sigma_j$, where weights $w_j$ are normalized magnitudes of $d_j$ and the sign depends on whether the neuron is popularity-aligned or unpopularity-aligned. Together these parts turn the black-box embedding into a list of identifiable, adjustable popularity signals.","core_discovery":"On its own terms, the paper establishes that the popularity bias of SASRec can be localized to a small set of interpretable neurons in an SAE trained to reconstruct the model's final-layer user embeddings. After replacing each user's items with random popular items or random unpopular items and recording hidden-layer activations, the authors compute a standardized effect size per neuron; large positive values mean the neuron fires more for popular profiles, large negative values mean it fires more for unpopular ones. PopSteer then adjusts those activations by a normalized, magnitude-weighted shift, boosting $d<0$ neurons and suppressing $d>0$ neurons. On ML-1M and Last.fm, this produces a better fairness-accuracy frontier than reweighting and reranking baselines, and the ablation shows that the same accuracy budget spent on random Gaussian noise does not produce the fairness gains. A direct neuron-deactivation experiment confirms the causal direction: switching off popularity-positive neurons lowers the Gini index, while switching off unpopularity-positive neurons raises it.","pith_inferences":["The same probe-and-steer recipe should transfer to other biases that can be expressed as synthetic user profiles, such as recency, category, or provider exposure, by swapping the popular and unpopular item sets; the paper only demonstrates popularity.","If popular and unpopular items differ systematically in content features such as genre, language, or release period, the synthetic replacement may let those confounds leak into the effect-size scores; a natural robustness check is to repeat the probe with popular and unpopular items matched on item metadata.","Because the SAE is trained on final-layer embeddings rather than on raw item text, the neurons it labels as popularity-related may actually encode whatever correlates with popularity in the embedding space; comparing steering outcomes across SAE seeds or sparsity levels would test how stable the attribution is."],"forward_implications":["PopSteer can be applied after training, without retraining or changing the base recommender, so it fits into deployed systems as a post-hoc fairness layer.","The SAE's per-neuron popularity scores give practitioners an audit trail: a system maintainer can see which internal features carry the bias and how strongly.","Tuning the steering strength $\\alpha$ and the number of adjusted neurons $N$ traces a fairness-accuracy frontier, allowing operators to choose how much accuracy they trade for exposure fairness.","The ablation implies targeted steering is doing the work: random perturbation of the same number of neurons at the same nDCG budget does not improve fairness.","The neuron-deactivation experiment turns interpretability into a causal check: the Gini index moves in the expected direction when identified neurons are switched off, confirming the identified neurons are causally linked to popularity bias."],"supporting_citations":[{"why":"Supplies the sparse autoencoder design, including top-K activation and the auxiliary dead-neuron loss, that PopSteer is built on.","marker":"[8]"},{"why":"Defines the SASRec sequential recommendation model used as the backbone that PopSteer attaches to and steers.","marker":"[16]"},{"why":"Defines the standardized mean-difference effect size used to score each neuron's popularity alignment.","marker":"[6]"},{"why":"Provides evidence that sparse autoencoders can isolate monosemantic, interpretable features, motivating neuron-level interpretation.","marker":"[3]"},{"why":"Defines the popular and unpopular head and tail item sets used to construct the synthetic probing profiles.","marker":"[1]"},{"why":"The MovieLens 1M dataset used for the first evaluation.","marker":"[12]"},{"why":"The Last.fm dataset used for the second evaluation.","marker":"[24]"},{"why":"Inverse Popularity Ranking, the reweighting baseline whose fairness-accuracy trade-off PopSteer is compared against.","marker":"[30]"}],"fun_headline_variants":["Steering bias neurons: fairer lists with minimal accuracy loss","Sparse autoencoder reveals exact neurons behind popularity bias","Pinpoint the bias neurons, nudge them: fairer recommendations","Interpretable fix for popularity bias: neuron steering wins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that replacing every item in a user's history with a random popular or unpopular item isolates the popularity signal while leaving everything else about the user unchanged; if popularity correlates with item content, category, or timing, the selected neurons may be responding to those confounds rather than to popularity itself.","fun_headline_variants_meta":{"raw":{"variants":["Steering bias neurons: fairer lists with minimal accuracy loss","Sparse autoencoder reveals exact neurons behind popularity bias","Pinpoint the bias neurons, nudge them: fairer recommendations","Interpretable fix for popularity bias: neuron steering wins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000464,"raw_usage":{"total_tokens":2308,"prompt_tokens":927,"completion_tokens":1381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":1313}},"tokens_in":543,"tokens_out":1381,"duration_ms":11612,"temperature":1.0,"reasoning_tokens":1313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:05:38.146581+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic profiles in which popular and unpopular items are matched on genre, category, and release-era distributions before computing the per-neuron effect sizes; if the neurons identified as popularity-positive stop predicting the Gini-index improvement from steering, the attribution is driven by those confounds and not by popularity alone.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides evidence that sparse autoencoders can isolate monosemantic, interpretable features, motivating neuron-level interpretation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inverse Popularity Ranking, the reweighting baseline whose fairness-accuracy trade-off PopSteer is compared against."}],"review_version":1}