{"id":"09ef77d1-262a-41dd-87d0-ecdb714480b9","arxiv_id":"2605.15208","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"3-bit quantization induces new stereotypical biases in 6-21% of previously unbiased BBQ items across three LLMs, undetected by perplexity increases under 3%, with models declining in 'unknown' responses by 17.4%.","lead":"The paper finds that quantizing instruction-tuned LLMs to low precisions like 3 bits causes new stereotypical biases to emerge in 6-21% of previously unbiased BBQ benchmark items, with a clear dose-response pattern, while standard metrics like perplexity show little change. A smart generalist should read it to understand hidden fairness risks when deploying compressed AI models on edge devices or in production.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.3","headline":"Decline in 'unknown' answers may drive apparent stereotypical shifts via general abstention failure rather than specific bias emergence","rationale":"The reader's weakest assumption correctly flags the risk that observed changes reflect degradation artifacts rather than genuine bias. My concern narrows this to one concrete, testable mechanism (non-specific drop in abstention) that is directly implied by the abstract's own 17.4% statistic and would falsify the 'stereotypical behaviors' framing if the test fails. This keeps the verdict conditional because the empirical setup is otherwise controlled but the interpretation of what the shifts mean remains under-determined without the proposed distribution check.","tokens_in":1824,"tokens_out":405,"duration_ms":42410,"concrete_test":"For each model, tabulate per-item response changes from BF16 baseline to 3-bit: compute the count of shifts to stereotypical vs. anti-stereotypical answers among items that previously selected 'unknown'. Apply a paired sign test or McNemar's test across the 12k items; if the excess of stereotypical shifts is not statistically significant (p>0.05) or smaller than 2 percentage points, the specific bias-emergence interpretation does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim requires that 6-21% of items develop new stereotypical behaviors specifically due to quantization undoing alignment. The abstract reports a 17.4% drop in 'unknown' selections at 3-bit. In BBQ ambiguous contexts, defaulting away from 'unknown' to one of the two named options would increase both stereotypical and anti-stereotypical answers roughly equally if the choice is non-preferential. The reported dose-response and item-level percentages would then reflect capability degradation in following the 'say unknown if ambiguous' instruction rather than targeted stereotype amplification. This distinction is load-bearing because the headline safety conclusion ('fairness-critical failures' missed by perplexity) depends on the shifts being bias-specific.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.3","summary":"The paper claims that post-training quantization of instruction-tuned LLMs (Qwen2.5-7B, Mistral-7B, Phi-3.5-mini) at precisions from BF16 to 3-bit induces new stereotypical biases on the BBQ benchmark. Across 12,148 items and 5 seeds (911,100 inferences), 6-21% of previously unbiased items develop stereotypical responses at 3-bit following a logistic-regression-confirmed dose-response, accompanied by a 17.4% drop in 'unknown' selections; these shifts are invisible to perplexity, which rises <0.5% at 8-bit and <3% at 4-bit.","tokens_in":1948,"tokens_out":647,"duration_ms":37746,"significance":"If the central interpretation holds, the work is significant for demonstrating that standard aggregate metrics miss fairness-critical degradation under compression. Strengths include the multi-model, multi-precision design, large item count with statistical controls, explicit dose-response modeling via logistic regression, and the scale of controlled comparisons that allow distinguishing gradual from threshold effects.","major_comments":[{"comment":"§4.2 (item-level shift analysis): The reported 6-21% emergence of new stereotypical behaviors is not accompanied by a breakdown showing whether anti-stereotypical answers increase at a comparable rate. Given the 17.4% decline in 'unknown' selections, a non-preferential default to one of the two named options would be expected to raise both stereotypical and anti-stereotypical counts roughly equally; without this disaggregation or a control for uniform abstention failure, the attribution to specific bias emergence (rather than general instruction-following degradation) is not yet established and directly affects the 'fairness-critical failures' conclusion.","section":"§4.2"},{"comment":"Methods §2.1 and Appendix A: The quantization implementation details (library, calibration dataset, group size, and any post-quantization fine-tuning) are described at a high level only. Because the central claim concerns threshold-dependent safety failures at 4-bit and 3-bit, the absence of these parameters prevents independent verification that the observed item-level changes are reproducible rather than artifacts of a particular quantization recipe.","section":"Methods §2.1"}],"minor_comments":[{"comment":"Abstract: The 6-21% range should be disaggregated by model or reported with per-model confidence intervals to allow readers to assess consistency.","section":"Abstract"},{"comment":"Figure 2 (dose-response plots): Add error bars or shaded regions reflecting the 5 random seeds so that the logistic regression fit can be visually assessed for robustness.","section":"Figure 2"},{"comment":"Table 1: Clarify whether the 'previously unbiased' baseline is computed per model or pooled across models, as this affects the denominator for the 6-21% statistic.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive and detailed feedback. The comments raise important points about distinguishing bias emergence from general degradation and ensuring reproducibility. We address each major comment below and indicate the revisions planned for the next version of the manuscript.","responses":[{"response":"We agree that a disaggregation of anti-stereotypical responses is needed to strengthen the attribution to specific bias rather than uniform degradation in instruction following or abstention. Our logistic regression analysis models the dose-response specifically for stereotypical outputs, but we acknowledge the current presentation does not explicitly compare rates of increase across categories. In the revised manuscript we will add this breakdown to §4.2, reporting the relative changes in stereotypical, anti-stereotypical, and unknown selections across precision levels together with statistical tests for differential effects. This addition directly addresses the concern and supports the fairness-critical interpretation.","revision_made":"yes","referee_comment":"[§4.2] §4.2 (item-level shift analysis): The reported 6-21% emergence of new stereotypical behaviors is not accompanied by a breakdown showing whether anti-stereotypical answers increase at a comparable rate. Given the 17.4% decline in 'unknown' selections, a non-preferential default to one of the two named options would be expected to raise both stereotypical and anti-stereotypical counts roughly equally; without this disaggregation or a control for uniform abstention failure, the attribution to specific bias emergence (rather than general instruction-following degradation) is not yet established and directly affects the 'fairness-critical failures' conclusion."},{"response":"The referee is correct that the current description is insufficient for full reproducibility. In the revised manuscript we will expand §2.1 and Appendix A to specify the quantization library and version, the exact calibration dataset, the group size parameter, and an explicit statement that no post-quantization fine-tuning was applied. These details will allow independent replication of the observed threshold effects at 4-bit and 3-bit.","revision_made":"yes","referee_comment":"[Methods §2.1] Methods §2.1 and Appendix A: The quantization implementation details (library, calibration dataset, group size, and any post-quantization fine-tuning) are described at a high level only. Because the central claim concerns threshold-dependent safety failures at 4-bit and 3-bit, the absence of these parameters prevents independent verification that the observed item-level changes are reproducible rather than artifacts of a particular quantization recipe."}],"tokens_in":1553,"tokens_out":535,"duration_ms":67299,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that 3-bit quantization leads to new stereotypical answers on 6-21% of BBQ items across models, with a clear dose-response, while perplexity stays almost flat. This suggests current checks miss fairness issues in compressed models. The work stands out for its scale and design. Three instruction-tuned models, five precision levels from BF16 to 3-bit, 12k+ items, five seeds, and logistic regression to track the pattern. Item-level analysis beats the usual aggregate bias scores, and the multi-model approach makes the finding more general. The soft spots are around interpretation. The reported 17.4% drop in 'unknown' answers could explain some shifts if models just pick an option more often in ambiguous cases. If the new answers split evenly between stereotypical and anti-stereotypical, it points to general instruction-following degradation rather than bias emergence specifically. The paper claims stereotypical behaviors, so they probably checked the polarity, but I'd want to see the exact counts to be sure. Methods details are a bit thin in the abstract too, like how quantization was applied exactly or full controls for other factors. This paper suits people focused on LLM safety and efficient deployment. Readers evaluating compression techniques or fairness benchmarks will get practical takeaways from the controlled comparisons. It has enough structure and data to merit serious referee time. I would recommend sending it to peer review, asking for more on response distributions and robustness checks.","headline":"Low-bit quantization introduces new stereotypical responses on BBQ items that perplexity misses, but the drop in 'unknown' answers could reflect general degradation more than targeted bias.","tokens_in":2453,"tokens_out":367,"would_cite":false,"duration_ms":40300,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":{"model":"grok-4.3","evidence":[{"relation":"unclear","rs_module":"IndisputableMonolith/Cost/FunctionalEquation.lean","rs_theorem":"washburn_uniqueness_aczel","paper_passage":"3-bit quantization causes 6-21% of previously unbiased items to develop new stereotypical behaviors, following a clear dose-response pattern confirmed via logistic regression"},{"relation":"unclear","rs_module":"IndisputableMonolith/Foundation/ArithmeticFromLogic.lean","rs_theorem":"LogicNat.induction","paper_passage":"models' willingness to select 'unknown' answers declines by 17.4%"}],"headline":"Empirical study of LLM quantization bias on BBQ benchmark; no structural overlap with RS forcing chain or cost functions","alignment":"orthogonal","rationale":"Paper's central machinery is controlled multi-precision empirical evaluation of stereotype reliance and unknown-selection rates under post-training quantization, using logistic regression and item-level transition analysis. This operates entirely in the domain of ML model compression and fairness benchmarks. RS framework derives spacetime, constants, and J-cost from a single distinction via modules such as AbsoluteFloorClosure, AlexanderDuality, ArithmeticFromLogic, and Cost.FunctionalEquation; the paper invokes none of these structures, makes no parameter-free constant derivations, and contains no cosh/J/φ-ladder reasoning.","tokens_in":47741,"confidence":"high","tokens_out":311,"duration_ms":14128,"cache_read_input_tokens":38528,"cache_creation_input_tokens":0},"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Quantization at 3 bits causes 6-21% of unbiased LLM items to develop new stereotypes while perplexity barely changes.","keywords":["quantization","bias emergence","large language models","model compression","fairness evaluation","BBQ benchmark","precision levels"],"falsifier":"Repeating the controlled runs and finding no statistically significant rise in stereotypical answers at 3-bit or 4-bit precision relative to BF16.","tokens_in":2693,"feed_emoji":"⚠️","tokens_out":601,"duration_ms":45169,"temperature":0.7,"pith_summary":"The paper tests whether post-training quantization preserves model behavior beyond standard quality measures. It tracks responses item by item on the BBQ benchmark as precision drops from BF16 to 3 bits in three instruction-tuned models. Results show a steady rise in new stereotypical answers that follows a dose-response curve, confirmed statistically, even though aggregate metrics like perplexity remain almost flat. The work demonstrates that fairness failures can appear at precision levels still considered safe by current evaluation standards. This gap matters because quantized models are widely deployed on edge devices where bias can affect real decisions.","feed_headline":"Quantization induces new biases in up to 21% of LLM responses at 3 bits","feed_subtitle":"Perplexity stays nearly flat while item-level stereotypes rise across three models and five precisions","key_machinery":"Item-level bias tracking on 12,148 BBQ items across five precision levels and three models, analyzed with logistic regression for dose-response.","core_discovery":"3-bit quantization causes 6-21% of previously unbiased BBQ items to develop new stereotypical behaviors following a clear dose-response pattern, with models' willingness to select unknown answers declining by 17.4%; these item-level shifts remain invisible to perplexity, which rises less than 0.5% at 8-bit and under 3% at 4-bit across models.","pith_inferences":["Compression pipelines for production use may need separate bias audits at each precision step.","Other compression methods such as pruning could produce similar hidden fairness shifts.","Post-quantization calibration focused on uncertainty and neutrality might reduce the observed bias increase."],"forward_implications":["Even 4-bit models already show new bias in 2.5-5.6% of items despite minimal perplexity shift.","Models become less willing to answer unknown, reducing by 17.4% at lowest precision.","Aggregate quality metrics systematically miss fairness degradation during compression.","Deployment of quantized models requires explicit bias testing before release."],"fun_headline_variants":["3-bit quantization forms new biases in 6-21% of items","Perplexity misses emerging biases in quantized LLMs","Dose-response bias appears across compression levels","Unknown answers drop 17% in 3-bit LLMs","Quantization undoes alignment on bias benchmarks"],"cache_read_input_tokens":64,"weakest_assumption_plain":"Changes in responses to BBQ items at lower bit widths reflect genuine emergence of stereotypical bias rather than random variation, model degradation, or evaluation noise.","fun_headline_variants_meta":{"raw":{"variants":["3-bit quantization forms new biases in 6-21% of items","Perplexity misses emerging biases in quantized LLMs","Dose-response bias appears across compression levels","Unknown answers drop 17% in 3-bit LLMs","Quantization undoes alignment on bias benchmarks"]},"model":"grok-4.3","cost_usd":0.008579,"raw_usage":{"total_tokens":3822,"prompt_tokens":727,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":85790500,"prompt_tokens_details":{"text_tokens":727,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":3020,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":727,"tokens_out":75,"duration_ms":34637,"temperature":1.0,"reasoning_tokens":3020,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-19T17:50:48.954164+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Repeating the controlled runs and finding no statistically significant rise in stereotypical answers at 3-bit or 4-bit precision relative to BF16.","supporting_citations":[],"review_version":1}