{"id":"cf07e65d-af2c-4d33-903c-906888c90031","arxiv_id":"2412.00323","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Testing two crowd-sourcing prompts on GPT-3.5 and GPT-4 shows that an awareness reminder (AwaRe) reduces several cognitive biases in LLM evaluation, while a social-projection prompt (SoPro) is mostly ineffective or harmful.","lead":"The paper surveys cognitive biases in large language models and tests whether two human crowd-sourcing debiasing prompts, SoPro and AwaRe, reduce six measured biases in GPT-3.5 and GPT-4. The results suggest AwaRe reduces some biases, especially bandwagon effects, while SoPro can make them worse, giving LLM users a cheap prompt-level mitigation option.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The abstract claims AwaRe makes LLMs 'more rational', but its evidence is only CoBBLEr score drops; because the AwaRe prompt names the exact bias being tested, the drops may be instruction-following rather than debiasing, so a control-prompt condition is required.","rationale":"The reader's weakest assumption identifies the validity of CoBBLEr consistency scores as the load-bearing point; my concern agrees with that direction but sharpens it into a specific alternative explanation: the AwaRe prompt names the target bias, so a score drop may reflect instruction-following or demand characteristics rather than a genuine increase in rational judgment. This does not contradict the reader's conditional verdict; it reinforces the need for control conditions and softened claims. I considered whether the lack of statistical significance is the more load-bearing issue, and it is real, but it is secondary because even with significance, the construct-validity confound would remain. The paper does have independent value: it surveys a wide literature, adapts two established crowdsourcing mitigation methods, and reports transparent tables for two models. However, the abstract's strong causal language is not supported by the design. Since the reader already recommended CONDITIONAL with requests for softened claims and significance testing, my analysis does not move the verdict; it adds a concrete control-prompt requirement to the conditions. I therefore set verdict_should_be to UNCHANGED and agreement_with_reader to partial, since I am refining the reader's assumption rather than identifying a wholly different one.","tokens_in":16664,"tokens_out":6917,"duration_ms":65168,"concrete_test":"Run the same 6-bias CoBBLEr evaluation with AwaRe replaced by control prompts that do not name the target bias: (i) 'Please answer carefully and rationally'; (ii) 'Please answer while being aware of syntax bias' (a nonexistent bias); (iii) for the bandwagon condition, 'Please ignore the statement about what percentage of people believe'. If any control reduces S_Band or the other bias scores by as much as the real AwaRe prompt, the observed mitigation is not specific to bias awareness and the abstraction to 'more rational' behavior is unsupported. Also report bootstrap 95% CIs for each Table 2 baseline-vs-AwaRe delta to determine which differences exceed noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AwaRe enables LLMs to mitigate biases and make more rational responses rests entirely on decreases in CoBBLEr consistency scores (Tables 2 and 3). No evidence is given that such decreases track rationality rather than compliance. The AwaRe prompt (Sec. 3.2) explicitly names the bias under test and tells the model to be careful of it; in the bandwagon condition this effectively instructs the model to disregard the inserted majority statement. A model that simply follows the salience of this instruction will show a lower S_Band (Eq. 6) without any improvement in judgment. The same confound applies to each bias: only in the treatment condition is the manipulated cue labeled as a bias to avoid. There is no control prompt (e.g., generic careful answering, an irrelevant or nonexistent bias name, or an explicit instruction to ignore the majority statement) to separate genuine debiasing from instruction-following or acquiescence. Further, the conclusion is drawn from point estimates with no confidence intervals; even under the paper's own metric, AwaRe worsens or leaves unchanged several entries in Table 2 (e.g., GPT-3.5 verbosity 0.096 to 0.119; GPT-4 egocentric 0.025 to 0.027; GPT-4 compassion-fade components both worsen). The abstract's plural claim that AwaRe mitigates 'these biases' therefore overstates the observed pattern. The post-hoc relabeling of invalid attentional-bias responses (Table 3) adds another untested assumption: that format violations can be relabeled as valid without changing the bias measurement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper surveys cognitive biases reported in LLMs and proposes applying two crowdsourcing-inspired prompt interventions, SoPro (social projection) and AwaRe (awareness reminder), to mitigate six biases in GPT-3.5 and GPT-4. The authors evaluate the interventions using the CoBBLEr benchmark, reporting bias scores before and after each intervention. They find that SoPro is largely ineffective or harmful (e.g., it worsens the bandwagon effect), while AwaRe appears to reduce several bias scores, most notably the bandwagon effect. The paper concludes that AwaRe enables LLMs to mitigate the effect of these biases and make more rational responses. The survey component (Table 1) is a useful systematization, but the experimental evidence for the mitigation claim is undermined by the lack of statistical testing, the instruction-following confound inherent in the AwaRe prompt, and a post-hoc relabeling of invalid responses in the attentional-bias analysis.","tokens_in":17026,"tokens_out":2321,"duration_ms":21112,"significance":"If the mitigation result held, it would offer a simple, low-cost, and bias-agnostic prompt intervention for improving LLM evaluator rationality, which is valuable for LLM-as-judge applications and for alignment. The paper's survey of existing literature is a genuine contribution, organizing more than 30 studies by bias type and mitigation. The experimental setup is transparent: the authors use a public benchmark, fix temperature and seed, and report per-condition scores. However, the central claim that AwaRe 'enables LLMs to mitigate the effect of these biases and make more rational responses' (Abstract) is not adequately supported by the evidence presented, for the reasons detailed in the major comments. The observed bandwagon reduction is large and consistent across both models, but the absence of control conditions and significance testing leaves the interpretation open to a simpler explanation: the model is following the explicit instruction to disregard the majority cue.","major_comments":[{"comment":"The AwaRe prompt directly names the target bias and instructs the model to be careful about it (e.g., 'Please answer the following question while being aware of order bias'). Consequently, a decrease in a bias score may reflect instruction-following rather than genuine debiasing: when the prompt tells the model that '80% of people believe...' is a bias to avoid, the model may simply discount the majority statement. This confound is present for every bias in Table 2, and it is load-bearing because the abstract's claim rests entirely on these score deltas. The authors should add control conditions, such as a generic 'please answer carefully' prompt, a prompt naming an irrelevant bias, or a prompt that explicitly tells the model to ignore the inserted cue without labeling it as a bias. Without such controls, the observed improvement cannot be attributed to bias mitigation rather than to the model's compliance with the stated instruction.","section":"§3.2, §4.3, Table 2"},{"comment":"The conclusion that AwaRe mitigates 'these biases' is drawn from point estimates without any confidence intervals, significance tests, or multiple-comparison correction. Several entries in Table 2 show no improvement or even worsening under AwaRe: GPT-3.5 verbosity worsens from 0.096 to 0.119; GPT-4 egocentric bias changes negligibly (0.025 to 0.027); and GPT-4 compassion-fade components move in opposite directions (SComp1 0.062 to 0.064, SComp2 0.073 to 0.076). The bandwagon effect and a few other cells improve, but the abstract's plural claim that AwaRe mitigates the effect of 'these biases' overstates the observed pattern. The paper should restrict its claim to the biases with consistent improvements, or add statistical analysis and discuss the mixed results.","section":"§4.3, Table 2"},{"comment":"The attentional-bias analysis relabels responses that violate the output format as 'valid' after the fact, with no justification that this correction is neutral with respect to the bias measurement. Table 3 shows that before correction, GPT-3.5 baseline has only 71.3% valid responses; after relabeling, the reported SAttn changes from 0.001 to 0.004. This post-hoc procedure changes the quantity being measured and could systematically alter the bias score if format violations correlate with the manipulated irrelevant information. The authors should either report the uncorrected results as the primary outcome, or provide evidence that the relabeling does not bias the SAttn estimates.","section":"§4.3, Table 3"}],"minor_comments":[{"comment":"There are several typographical artifacts in the manuscript, including 'evaluation se/t_ting' (Section 4.2) and 'A/t_tentional bias' in section headings and Table 3. These should be corrected.","section":"Throughout"},{"comment":"The description of the CoBBLEr evaluation says the LLM performs 12,000 evaluations per bias, but the relationship between the 50 questions, 16 responses, and the number of pairwise comparisons is not fully spelled out. A short derivation or reference to the original CoBBLEr protocol would make the experimental load and score denominators clearer.","section":"§4.1"},{"comment":"The phrase 'more rational responses' is used in the Abstract and Discussion, but the evaluation only measures consistency of pairwise preferences under manipulation. A decrease in a consistency-based score is not the same as an increase in rationality; the authors should either define rationality in terms of consistency or soften the claim to 'more consistent' responses.","section":"§1, §4.4"},{"comment":"The prompt templates in Appendix B are useful, but the bandwagon template places the majority statement after the two system responses and before the format instruction; the exact position relative to the baseline template is not discussed, even though prompt ordering can affect LLM behavior. This should be noted or controlled.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful survey and a clear, reproducible experimental setup, but the central mitigation claim is stronger than the evidence. The main issue is not the absence of significance tests per se, but that the AwaRe condition is confounded with explicit instruction-following, which is a conceptual problem rather than a mere statistical one. The bandwagon result is suggestive and worth following up, but the paper as written overstates the generality of the finding. A major revision that adds control prompts and restricts claims to the biases with consistent improvements would make the contribution solid."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWorth a look if you work on LLM-as-judge or prompt-based debiasing, but the headline claim overshoots the evidence. What's actually new: applying two crowdsourcing debiasing prompts (SoPro and AwaRe) to GPT-3.5/GPT-4 on CoBBLEr, and the negative result that SoPro—answering as you think the majority would—worsens the bandwagon effect. That is a clean, mechanistically sensible finding. The survey table is a useful map of the bias-mitigation literature even if it is mostly reorganization.\n\nThe paper does several things right. The experimental setup follows CoBBLEr's protocol, prompts are in the appendix, and temperature and seed are fixed. The claims about SoPro's failure are appropriately limited. The AwaRe improvements on bandwagon are large and consistent across both models (0.524 to 0.260 on GPT-3.5, 0.214 to 0.142 on GPT-4).\n\nNow the soft spots, which are serious. The AwaRe prompt literally names the bias under test: 'Please answer the following question while being aware of order bias.' In the bandwagon condition it tells the model to ignore the majority statement. A model following instructions will lower the consistency-based bias score without any change in judgment quality. There is no control prompt—neither a generic 'answer carefully' nor a placebo bias name—so you cannot separate instruction-following from debiasing. The abstract's plural claim that AwaRe mitigates 'these biases' also overstates the table: several entries get worse or stay flat (GPT-3.5 verbosity, GPT-4 egocentric, GPT-4 compassion fade). No error bars or significance tests, and the post-hoc relabeling of invalid responses for attentional bias adds another untested assumption.\n\nNone of this sinks the paper. The SoPro-negative result is worth publishing, and AwaRe deserves a follow-up with proper controls. But the current abstract needs softening and the statistics need to catch up with the claims.\n\nBottom line: send it to review—it is a legitimate empirical contribution—but expect revisions. For a reading group, it is a good case study in why prompt-based debiasing evaluations need control conditions.","headline":"Useful new negative result on SoPro plus a promising AwaRe effect, but the no-control prompt design means the main debiasing claim is instruction-following in disguise.","tokens_in":17538,"tokens_out":2072,"would_cite":true,"duration_ms":20089,"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":"A short awareness prompt makes GPT-3.5 and GPT-4 measurably less biased, the paper claims.","keywords":["Large Language Models","cognitive bias","debiasing","bandwagon effect","order bias","prompt engineering","CoBBLEr","AwaRe"],"falsifier":"If running AwaRe with a made-up bias name on the same benchmark reduces scores as much as the real AwaRe does, the measured mitigation is generic instruction-following rather than bias-specific awareness, which would undercut the paper's interpretation. A second decisive check is to have humans rank the same response pairs as rational or irrational and see whether AwaRe's score reductions align with improved human-rated rationality.","tokens_in":16496,"feed_emoji":"🧠","tokens_out":4876,"duration_ms":39248,"temperature":0.7,"pith_summary":"The paper claims that a two-sentence prompt reminder can make large language models less susceptible to six cognitive biases, most clearly the bandwagon effect. The authors survey prior evidence that LLMs inherit human-like biases from training data, then adapt two human-focused debiasing instructions from crowdsourcing into LLM prompts. In experiments on GPT-3.5 and GPT-4 using the CoBBLEr bias benchmark, the AwaRe (awareness reminder) prompt lowered bias scores, while the SoPro (social projection) prompt often failed or amplified bias. The claim matters because the method requires no retraining, no repeated sampling, and no lengthy outputs, only a short instruction naming the bias.","feed_headline":"Awareness prompt halves GPT-3.5's bandwagon bias","feed_subtitle":"A short 'be aware of this bias' reminder outperforms social-projection prompting on six cognitive biases in GPT-3.5 and GPT-4.","key_machinery":"Two prompt-level interventions are the mechanism. AwaRe prepends a sentence naming the bias and telling the model to be careful of it, e.g., 'Please answer the following question while being aware of order bias.' SoPro prepends instructions to answer as the majority of people would. Their effectiveness is measured with CoBBLEr, a benchmark that computes scores such as S_Band and S_Verb by checking whether an LLM changes its preferred answer when the order, names, majority signals, or distracting content in a prompt are flipped; a score of zero means no bias, and scores above random (0.25 for most biases) mean the bias is influencing choices.","core_discovery":"On the paper's own terms, the central discovery is that telling an LLM to be aware of a specific bias before it answers — the AwaRe prompt — reduces the measured influence of that bias on its pairwise quality evaluations, whereas asking it to answer as it believes the majority would (SoPro) does not and can make conformity biases worse. Using the CoBBLEr benchmark, which scores bias by seeing whether an LLM changes its preference when inputs are permuted, the authors measure six biases. AwaRe lowered the bandwagon-effect score from 0.524 to 0.260 for GPT-3.5 and from 0.214 to 0.142 for GPT-4, with smaller gains on several other biases, while SoPro raised the bandwagon score to 0.955 for GPT-3.5 and 0.255 for GPT-4. The authors conclude that awareness prompting nudges LLMs toward more rational responses.","pith_inferences":["A likely confound not settled by the paper is that AwaRe's gain is generic instruction-following: a prompt that names any bogus bias and asks the model to be careful might lower consistency scores just as well, a test that can be run with a fabricated bias label on the same benchmark.","The consistency-based CoBBLEr scores treat 'changing one's mind when the prompt is permuted' as bias, but a model could be consistent the wrong way; no human-rationality ground truth is given, so 'more rational' is an interpretation, not a proven fact.","An extension beyond the paper would test AwaRe on open-source models of different sizes and on human-validated rationality tasks to see whether the score drops correspond to genuinely better decisions.","AwaRe's reliance on naming the bias means it cannot address unknown biases; a future direction is automatic bias detection feeding the prompt."],"forward_implications":["If AwaRe's effect is real, any LLM answer can be made less biased by adding a short bias-naming instruction, with no fine-tuning.","SoPro should be avoided for conformity-related biases because it can amplify the bandwagon effect, as the GPT-3.5 score of 0.955 shows.","Newer models such as GPT-4 show lower baseline bias, suggesting bias mitigation and model capability improve together.","Bias scores for some biases such as attentional bias are near zero at baseline, so AwaRe's main measurable wins are on bandwagon, verbosity, and egocentric biases.","The method requires knowing in advance which bias will affect the model, since AwaRe must name the bias."],"supporting_citations":[{"why":"Supplies the CoBBLEr benchmark and the six bias scores used to measure mitigation.","marker":"[26]"},{"why":"Defines SoPro and AwaRe, the two crowdsourcing mitigation prompts adapted here.","marker":"[20]"},{"why":"Provides the existing 'with reason' debiasing method used as the comparison baseline.","marker":"[52]"},{"why":"Grounds the claim that crowdsourcing bias-mitigation techniques apply to cognitive biases generally.","marker":"[17]"}],"fun_headline_variants":["AwaRe prompt halves GPT-3.5 bandwagon bias","Simple bias-awareness prompt curbs LLM conformity","Awareness outdoes social projection in debiasing LLMs","AwaRe bias cue beats SoPro for rational LLM answers","Telling LLMs 'be aware' cuts six cognitive biases"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that CoBBLEr's consistency scores—checking whether a model reverses its preference when the prompt is permuted—truly measure cognitive bias and that lower scores mean more rational behavior, with no human ground-truth answers to confirm this.","fun_headline_variants_meta":{"raw":{"variants":["AwaRe prompt halves GPT-3.5 bandwagon bias","Simple bias-awareness prompt curbs LLM conformity","Awareness outdoes social projection in debiasing LLMs","AwaRe bias cue beats SoPro for rational LLM answers","Telling LLMs 'be aware' cuts six cognitive biases"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1553,"prompt_tokens":952,"completion_tokens":601,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":514}},"tokens_in":568,"tokens_out":601,"duration_ms":5893,"temperature":1.0,"reasoning_tokens":514,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:29:11.684187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"If running AwaRe with a made-up bias name on the same benchmark reduces scores as much as the real AwaRe does, the measured mitigation is generic instruction-following rather than bias-specific awareness, which would undercut the paper's interpretation. A second decisive check is to have humans rank the same response pairs as rational or irrational and see whether AwaRe's score reductions align with improved human-rated rationality.","supporting_citations":[],"review_version":1}