Pith. sign in

REVIEW 4 major objections 5 minor 5 references

Meaningless is better: hashing bias-inducing words in LLM prompts improves performance in logical reasoning and statistical learning

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Replacing bias-triggering words with meaningless identifiers makes LLMs answer the Linda problem correctly and improves frequent-itemset recall, the paper reports.

desk verdict A useful but weakly controlled prompt-debiasing idea; the unique-identifier twist is real, but the main Linda experiment conflates hashing with prompt rewording and bespoke word selection. read the letter →

arxiv 2411.17304 v2 pith:53DQZRQA submitted 2024-11-26 cs.CL cs.AI

Milena Chadimová , Eduard Jurášek , Tomáš Kliegr This is my paper · ORCID
classification cs.CLcs.AI
keywords largelanguagemodelscognitivebiasesconjunctionfallacyhashingpromptengineeringdebiasingfrequentitemsetsrepresentativenessheuristic
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that replacing words likely to trigger stereotypes or memorized associations with meaningless, hash-like identifiers improves LLM accuracy on tasks where those associations cause errors. In the modified Linda problem, models never picked the correct non-conjunction answer in 80 baseline runs, but after hashing they answered correctly 14 or 13 times out of 80 in the two prompt variants; in frequent-itemset counting, hashed tables raised recall for both tested models, with the largest gain for Llama-3.1-405b. A tabular version of the Linda problem also improved with hashing, and the method matched or beat chain-of-thought prompting on several comparisons without requiring extra inference compute. The authors position hashing as a prompt-level, model-agnostic debiasing step that reduces reliance on external knowledge, while noting that hallucination rates were not consistently reduced.

What carries the argument

Hash-like identifiers: each potentially bias-inducing word or phrase is replaced by a fixed meaningless token (for example, reading becomes rfg5a) that can still be referenced later in the prompt. The mechanism is that the model must reason about relations between opaque placeholders rather than about 'reading' and 'colorful coat', so its answers are driven by the prompt's structure and stated relationships instead of by semantic priors from training data. Because the identifiers are unique and repeatable, the method differs from ordinary masking and allows the prompt to state possible links between hidden elements without revealing their identities.

What would settle it

Hash a different set of content words on the same adapted Linda prompt—for instance, only the adjectives 'long' and 'colorful', or every noun except the target roles—and compare fallacy rates against the original baseline. If any content-word hashing produces the same near-elimination of the fallacy, the effect is generic; if only the paper's exact word list works, the result is an artifact of selection. A stronger test would choose the words on a development set of models and prompts, freeze the list, and apply it to held-out models and prompts.

Watch

Extended reading notes

Core claim

The central discovery is that a pure lexical transformation—replacing selected content words by unique opaque identifiers such as cdf14 while keeping syntax and cross-references intact—makes LLMs less likely to answer from learned stereotypes and more likely to answer from the logical structure of the prompt. In the conjunction-fallacy experiment, the original adapted Linda prompt produced the wrong conjunctive answer in 80 of 80 runs across Gemini, GPT-3.5, GPT-4, and Llama 2, while the hashed prompts yielded 14 and 13 correct answers out of 80, with Fisher's exact test p-values below 0.00001 and large Cramér's V effect sizes. In the itemset task, hashing increased true itemsets found from 213 to 225 out of 235 for GPT-4o and from 194 to 230 out of 235 for Llama-3.1-405b, both statistically significant by chi-square tests. The authors interpret this as disrupting the representativeness heuristic and preventing the model from injecting pretrained world knowledge into tasks where the prompt's artificial data should be the only source of truth.

Load-bearing premise

The success depends on the manually selected list of bias-inducing words being a stable, model-independent set; if the words were chosen because they made the specific test models fail on the specific prompts, the reported improvement is not evidence that hashing works on new prompts or models.

Editorial extensions

If this is right

  • On the adapted Linda problem, hashing raised correct non-conjunction answers from 0 out of 80 to 14 or 13 out of 80, with p-values below 0.00001 and Cramér's V around 0.47–0.49.
  • On frequent-itemset extraction, hashed data improved recall for GPT-4o (from 213 to 225 of 235 true itemsets) and for Llama-3.1-405b (from 194 to 230), with statistically significant chi-square tests in both cases.
  • In the tabular CSV version of the Linda problem, hashing without added relationship descriptions raised correct answers across models with p = 0.0000535 and Cramér's V = 0.404, including 10 out of 10 for Llama-3.1-405B.
  • Hashing was comparable to or better than chain-of-thought prompting on several GPT-family comparisons, and unlike CoT it adds no extra inference cost.
  • Hallucination reduction was inconsistent: GPT-4o hallucinated at similar rates across conditions, while Llama-3.1-405b hallucinated more in the hashed condition.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: because the GPT-4o itemset effect was statistically significant but small (Cramér's V ≈ 0.09), the practical benefit of hashing for counting tasks may be modest even where it is real, and the larger Llama recall jump suggests the gain scales with how strongly a model leans on pretrained associations.
  • Editorial inference: hashing hides identity while preserving co-reference and syntax, so combining it with chain-of-thought could suppress both associative bias and shallow reasoning; the paper names this combination as future work, and a direct head-to-head test would be a natural next experiment.
  • Editorial inference: automating the selection of words to hash would turn the method from a hand-crafted intervention into a deployable preprocessor, and the authors' manual selection is the main obstacle to scaling the approach.
  • Editorial inference: if the effect replicates on new prompts and models, hashing offers a cheap inference-time debiasing lever for human-in-the-loop systems where retraining is impractical, at the cost of hiding information that a user might want to inspect.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes 'hashing': replacing potentially bias-inducing words or data values in prompts with unique meaningless identifiers, and reports four experiments on the Linda conjunction-fallacy task (free-text and tabular variants), frequent-itemset counting, and a comparison with chain-of-thought (CoT) models. The paper claims statistically significant improvements in all tested scenarios, while noting that the effects vary across model families and that hallucination rates are inconsistently reduced. Prompts and model outputs are available in a public repository.

Significance. If the claimed effect is real, prompt-level hashing would be a lightweight, model-agnostic debiasing intervention with practical value, and the comparison with CoT is timely and useful. The paper's strengths include the clean construction in Experiment 2, where CSV-correct, CSV-wrong, and CSV-hashed datasets were designed to preserve the frequent-itemset ground truth; evaluation of recall and precision against Apriori; coverage of several model families; and a public repository of prompts and outputs. However, the central Experiment 1 evidence does not currently isolate the proposed mechanism, and the word-selection procedure introduces a test-set fitting component. These issues are load-bearing for the paper's main claim, so the manuscript requires substantial revision.

major comments (4)
  1. [§3.1.1, Figs. 5–6] The hashed prompt is not the original prompt with selected words replaced. It inserts an explicit instruction ('Note that in the text below, specific information was masked behind anonymous identifiers such as X and cdf14'), rewrites the person description in ungrammatical placeholder syntax ('Imagine X with a cdf14 and a a214s, sitting in a fg57 rfg5a'), and appends relationship hints. The validation condition in Fig. 7 controls only for the added neutral descriptions and relationship hints; it does not control for the masking note or for the placeholder syntax while keeping semantic content available. The observed improvement therefore cannot be attributed specifically to removing word meanings; it may stem from instructing the model that surface details are untrusted or from the syntactic disruption. A same-symbol masking baseline and a control that adds the masking note without hashing are needed.
  2. [§3.1.1 'Selection of possible bias-inducing words'] The words to hash were selected by inspecting the same models' incorrect outputs on the original prompt, and the Limitations section confirms this. The method therefore includes a bespoke, test-set-fitted mask, and the measured improvement is not evidence for hashing as a general method with a stable word set. The authors should either pre-specify the mask independently of the evaluated outputs, evaluate on held-out models or prompts, or demonstrate that the selected word set is stable across models. Without this, Experiment 1 conflates the hashing mechanism with the selection procedure.
  3. [§3.1.2, Tables 3–4 and Table 6] The pooled Fisher test (p < 0.00001) aggregates 20 or 10 iterations per model and is driven by large shifts in some models, but per-model results are highly heterogeneous: Llama 2 70B remains at 0 correct in both hashed variants, GPT-4 improves to 10/10 with added descriptions but drops to 1/10 without, and Gemini shows the opposite pattern. The abstract's claim of 'significant improvements in all tested scenarios' overstates the evidence. Per-model tests should be reported and the consistency of the effect discussed; the pooled N is small and the prompts are not identical across conditions.
  4. [§3.2.2, Tables 8 and 10] The chi-square tests treat individual itemsets (e.g., 235 itemsets per condition) as independent observations, although itemsets within one prompt or run are highly correlated. This inflates the effective sample size and can make small differences significant, as in GPT-4o's p = 0.043 with Cramér's V = 0.093. The analysis should be performed at the run level (25 runs per condition) or with a mixed-effects model that accounts for prompt-level clustering.
minor comments (5)
  1. [Abstract] The abstract states '490 prompts' in one version and '680 prompts' in the full manuscript; these numbers should be aligned.
  2. [§3.2.2] The text says 'In the Llama 2 model, a χ² test was also conducted' but the experiment uses Llama 3.1-405b; this appears to be a typo.
  3. [§3.4] The model name 'ChatGPT-3o-mini' should be 'ChatGPT-o3-mini'; also, the comparison text inconsistently refers to 'ChatGPT-4' and 'ChatGPT-4o'.
  4. [Tables 3 and 4] The option 'X is b321 who 4l5i' is ungrammatical; the intended reading is presumably 'X is b321 who likes to 4l5i' or 'X is b321 who 4l5i' should be otherwise clarified.
  5. [Figures 12 and 13] The figures would benefit from explicit axis labels and a note that hallucination counts are shown separately from found itemsets.

Circularity Check

1 steps flagged · score 6.0 of 10

The hashed-word set in the Linda experiments is selected from the models' original-prompt errors, so the reported improvement is partly a fitted-input effect; Experiment 2 provides independent but partial support.

  1. fitted input called prediction [Section 3.1.1, 'Selection of possible bias-inducing words'; reiterated in Section 4, Limitations]
    "To determine which of these words should be hashed, we analysed the models' responses to the original (non-hashed) prompt. Words that influenced the models' decision-making when incorrect answers were produced were considered for hashing. ... In the experiments, we identified the biased words manually, as described in Experiment 1, based on the models’ outputs on the non-hashed variants of the tasks."

    The treatment variable (which words are replaced by hashes) is fitted to the exact baseline being evaluated: the models' incorrect outputs on the original prompt. The headline improvement in Table 6 compares that original baseline (0/80 correct) with hashed variants (13/40 and 14/40 correct) in which the mask was chosen to remove the words implicated in those same incorrect outputs. The gain is therefore partly a consequence of selecting the mask to target observed failure modes, not a clean test of a fixed, model-independent hashing rule. The circularity is partial because Experiment 2 hashes all data values without a selection step, and Experiment 3 transfers the fitted mask to models that were not all used in the original selection, giving some out-of-sample content.

full rationale

The main circular step is confined to Experiments 1 and 3: the set of hashed words is derived from the models' responses to the original non-hashed prompt, and the same original condition is then used as the baseline for measuring improvement. Section 3.1.1 states this explicitly, and Section 4 reiterates that the biased words were identified manually based on those outputs. Thus the central Linda-based demonstration reduces in part to a fitted mask rather than to an independent test of hashing as a general method. This is not a mathematical equivalence, but it is a genuine fitted-input-called-prediction pattern: the experimental 'prediction' is closely tied to the data used to construct the intervention. Experiment 2 is substantially cleaner because the hashed condition replaces all data values with identifiers and is scored against an external Apriori reference solution; it also reports mixed hallucination effects, which are not selectively chosen. Experiment 4 compares against chain-of-thought and does not depend on the fitted mask for its main comparison. There is no load-bearing self-citation or uniqueness import from the authors' prior work; the reliance on Suri et al.'s prompt is external and disclosed. An additional non-circular validity concern exists in Experiment 1: the hashed prompt adds an explicit masking note and unusual placeholder syntax that the validation prompt does not fully control, so the active mechanism is not isolated. That concern weakens interpretation but does not itself constitute circularity. Overall, because one of the paper's central demonstrations is partly constructed by the word-selection procedure, while another experiment provides independent support, the appropriate circularity score is moderate.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The ledger contains no fitted numerical constants; the main fitted element is a categorical mask (which words to hash) chosen from model outputs, plus assumptions about semantic neutrality and about the reference solution. No new physical or theoretical entities are introduced.

free parameters (1)
  • Manual hashing mask for Experiment 1 = woman, long hair, colorful coat, corner of a café, reading, artist, likes to read (Table 1)
    The set of words and phrases to hash was chosen by inspecting which words led to incorrect answers on the original prompt, making it a categorical parameter fitted to the same data used for evaluation.
assumptions (4)
  • domain assumption LLM errors on the adapted Linda task arise mainly from the representativeness heuristic and from external pretraining knowledge, not from instruction comprehension or output-format effects.
    Sections 2.1 and 3.1 motivate the intervention this way; if the errors had another cause, removing semantic cues would not be the right lever.
  • domain assumption Replacing words with random hash-like identifiers preserves the logical structure of the task and does not itself bias the model.
    Section 3.2.1 constructs the CSV-hashed dataset by substitution and assumes the frequent-itemset solutions remain equivalent; the paper provides no control for token-frequency or formatting effects of the hashes.
  • standard math The Apriori algorithm outputs and the manual TP/FP classification of LLM responses are correct and unbiased.
    Section 3.2.2 uses algorithmically computed itemsets as ground truth; output-format or parsing errors could change the precision and recall counts.
  • domain assumption Removing the phrase 'Pretend to be an agent' from some hashed prompts does not invalidate comparisons with the original prompt.
    Section 3.1.1 states this removal for Copilot and Llama-3.1-70B after observing non-response; the paper does not quantify its effect on the comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Meaningless is better: hashing bias-inducing words in LLM prompts improves performance in logical reasoning and statistical learning." pith.science (2026). https://pith.science/paper/53DQZRQA

@misc{pith2026241117304,
  author       = {Pith},
  title        = {Pith review of: Meaningless is better: hashing bias-inducing words in LLM prompts improves performance in logical reasoning and statistical learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/53DQZRQA}},
  note         = {Machine review of arXiv:2411.17304}
}
read the original abstract

This paper introduces a novel method, referred to as "hashing", which involves masking potentially bias-inducing words in large language models (LLMs) with hash-like meaningless identifiers to reduce cognitive biases and reliance on external knowledge. The method was tested across three sets of experiments involving a total of 490 prompts. Statistical analysis using chi-square tests showed significant improvements in all tested scenarios, which covered LLama, ChatGPT, Copilot, Gemini and Mixtral models. In the first experiment, hashing decreased the fallacy rate in a modified version of the "Linda" problem aimed at evaluating susceptibility to cognitive biases. In the second experiment, it improved LLM results on the frequent itemset extraction task. In the third experiment, we found hashing is also effective when the Linda problem is presented in a tabular format rather than text, indicating that the technique works across various input representations. Overall, the method was shown to improve bias reduction and incorporation of external knowledge. Despite bias reduction, hallucination rates were inconsistently reduced across types of LLM models. These findings suggest that masking bias-inducing terms can improve LLM performance, although its effectiveness is model- and task-dependent.

Figures

Figures reproduced from arXiv: 2411.17304 by the authors.

Figure 12
Figure 12. Results of frequent itemset mining task of ChatGPT-4o showing the count of found itemsets and hallucinations. The more itemsets found and the fewer hallucinations, the better. Llama 3.1-405b results Llama 3.1-405b exhibited more variability in performance compared to GPT-4o. As shown in [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Results of frequent itemset mining task of Llama 3.1-405b showing the count of found itemsets and hallucinations. The more itemsets found and the fewer hallucinations, the better. These findings support our initial research questions for Experiment 2. They demonstrate that LLMs may rely on learned associations, which may impact results when input data contains potentially bias-inducing associations. The improved per… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 1 canonical work pages

  1. [1]

    Agrawal, R., Mannila, H., Srikant, R., Toivonen, H., & Verkamo, A. I. (1996). Fast discovery of association rules. In U. M. Fayyad, G. Piatetsky -Shapiro, P. Smyth, & R. Uthurusamy (Eds.), Advances in knowledge discovery and data mining (pp. 307–328). AAAI Press. Ayala-Chauvin, M., & Avilés -Castillo, F. (2024). Optimizing Natural Language Processing: A C...

    arXiv 1996

  2. [293]

    & Kalinli, O

    https://doi.org/10.1037/0033-295X.90.4.293 Vats, A., Liu, Z., Su, P., Paul, D., Ma, Y., Pang, Y., ... & Kalinli, O. (2024). Recovering from privacy-preserving masking with large language models. In ICASSP 2024 – 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 10771–10775). IEEE. Wang, P., Xiao, Z., Chen, H., & O...

    arXiv 2024

  3. [650]

    Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reasoning without instruction: frequency formats. Psychological review, 102(4),

    work page 1995

  4. [684]

    https://doi.org/10.1037/0033-295X.102.4.684 Gigerenzer, G., & Todd, P. M. (1999). Simple heuristics that make us smart. Oxford University Press. Hackmann, S., Mahmoudian, H., Steadman, M., & Schmidt, M. (2024). Word importance explains how prompts affect language model outputs [Preprint]. arXiv. https://arxiv.org/abs/2403.03028 Hagendorff, T., Fabi, S., &...

    arXiv 1999

  5. [1726]

    https://doi.org/10.3390/healthcare12171726 Liu, M., Okuhara, T., Dai, Z., Huang, W., Okada, H., Emi, F., & Kiuchi, T. (2024). Performance of advanced large language models (GPT-4o, GPT-4, Gemini 1.5 Pro, Claude 3 Opus) on Japanese Medical Licensing Examination: A comparative study [Preprint]. medRxiv. Macmillan-Scott, O., & Musolesi, M. (2024). (Ir) ratio...

    arXiv 2024

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.