acceptodds
Under review as a conference paper at ICLR 2027

Labels, sources and routes: Ten biases in language models

Abstract

A key hallmark of intelligence is the ability to make robust, reliable and stable decisions from data even under irreverent changes in structure or form of presentation of the data. Can language-model agents achieve this consistently? We propose a framework of ten operational biases: structural, positional, numerical, ordinal, statistical, informational, logical, semantic, cognitive, and self-referential. We investigate these using three experimental layers: static choices and rankings, attribution-based judgment and correction, and sequential navigation in a grid world. We analyze 381,258 responses and navigation episodes from 59 model endpoints across Cohere (5), Google (3), Meta (5), OpenAI (44), and xAI (2). Despite explicit instructions that labels were arbitrary, models exhibited preferences in 735 of 1,060 conditions (69.3%): 190 of 363 structural conditions (52.3%), 248 of 315 ordinal conditions (78.7%), and 297 of 382 semantic conditions (77.7%). These involved random-looking identifiers, ordered labels, and familiar words, respectively, suggesting that labels could bias choices independently of their information content. Positional effects were observed in 265 of 716 eligible first-option comparisons. In numerical probes, 38 of 43 eligible endpoints most frequently output 7 when asked for a random integer from 1 to 10. Logical-bias tests assessed whether models imposed assumptions about meaning, order, or quality that the instructions explicitly excluded. For self-referential bias, we held answer content fixed while varying description as the model’s own, another agent’s, or anonymous. At requested temperature 0.7, 29 of 41 high-coverage endpoints scored own-attributed answers more highly; the mean own-versus-other difference across all 41 was +0.12 on a seven-point scale. However, 27 of 41 endpoints corrected wrong answers equally often under both attributions. Self-attribution thus influenced ratings more consistently than willingness to revise. In grid navigation across 42 eligible endpoints, goal success dropped from 78.8% with good route advice to 3.8% with advice leading toward a lower-reward decoy. Nevertheless, models followed misleading advice more closely (87.2% versus 69.4%), consistent with route fixation or planning inertia. These results demonstrate model-specific profiles of bias and motivate evaluation of sensitivity to presentation, evidence, and plan revision, as well as task accuracy.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, LLM agents, behavioral biases, robustness evaluation, decision stability, presentation sensitivity, self-attribution, self-correction, planning inertia

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Xyy6OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates whether language-model decisions change with labels, source descriptions, and prior route advice. It organizes these behaviors into ten overlapping bias categories and evaluates static choices, answer evaluation and correction, and hidden-grid navigation across 59 model endpoints, collecting 381,258 response and episode records. The reported findings include label and position preferences, source-dependent ratings with mixed correction effects, and substantially lower navigation success under misleading advice. The intended contribution is a unified empirical characterization of these behaviors.

Critical strengths

The paper addresses a relevant evaluation problem. Examining sensitivity to presentation and advice across several decision settings yields useful descriptive observations about model behavior beyond isolated answer accuracy. The fixed-content attribution experiment includes a useful anonymous baseline. This helps distinguish favorable treatment of own-attributed answers from unfavorable treatment of other-attributed answers, as illustrated by the GPT-4o-mini results in Section 4.3.

Critical weaknesses

The scientific contribution is not sufficiently established. Related Work mentions prior research without explaining the added value over label-bias measurement [1], self-preference studies [2], and existing multi-bias evaluation such as CoBBLEr [3]. The ten categories combine input variations with summary measurements, but their additional explanatory or diagnostic value remains unclear. The central reliability interpretations are not adequately isolated. Under requested temperature zero, the evidence-free choice tasks permit any option, so unequal selection alone does not establish a decision failure. In navigation, agents lack global reward information and are not told that the decoy is suboptimal. The success gap therefore does not distinguish planning inertia from limited information and exploration difficulty. The Discussion acknowledges this ambiguity, but the relevant controls remain unevaluated. Limited evidence for generalization and attribution effects. The answer-evaluation experiments use six basic questions, and navigation uses two maps, limiting generalization. The supplementary code provides exact-body matching and model-level bootstrap analysis for attribution comparisons, which is helpful. However, the resulting matched-subset estimates, confidence intervals, and retained sample counts are not reported, leaving the robustness of the correction effects unclear. Incomplete reporting of robustness results. The supplement includes code for strict-validation and repair-exclusion analyses, but I could not locate the empirical results corresponding to Appendix E. The inconsistent headline counts between the abstract and main text also remain unresolved. These gaps make the reported findings difficult to verify.

Questions for authors

Which principal finding or evaluation capability is new relative to [1–3], and what additional insight does the ten-category framework provide? What evidence connects the measured label and position preferences to incorrect decisions when task-relevant evidence is available? Can no-advice and explicit decoy-warning or full-information controls distinguish planning inertia from information-access and exploration limitations? Under exact-body matching, including identical option order, how many attribution pairs remain, and do the correction effects persist with uncertainty estimates? Could the authors provide the Appendix E sensitivity results, including retained sample counts and principal effects after strict validation and repair exclusion, and reconcile the abstract and main-text statistics?

Reviewer_NwT6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies how LLMs output are effected when only presentation of input changes, while still responding to relevant evidence. The paper proposes 10 bias categories and evaluate them on 59 model endpoints. In the experiment, the authors measure the biases in three ways. Firstly, in static probes setting, models choose from four equivalent choices whose labels are declared arbitrary, and also select random integers. Secondly, in attribution probes setting, the authors change the phrasing of LLM answers as the model’s own generated, another agent’s, or anonymous, and then ask LLMs to rate or correct the modified answers. Thirdly, models search for the highest-reward cell while given a suggested route that leads either to the goal or to a lower-reward decoy. Models often prefer certain labels and positions among equavalant options, but they mostly follow visible scores. Describing an answer as the model’s own generated raises the rating slightly and does not consistently change correction. In the grid, target success is much lower under decoy advice than under target advice (3.8% vs. 78.8%), even though models follow decoy advice more closely. The paper concludes that should evaluate sensitivity to presentation, use of evidence, and plan revision separately from the task accuracy.

Critical strengths
  1. The research question is clear and pratical: how sensative the LLM outputs to the changes of presentation and changes of evidence. The topic is important for debiasing the LLMs and improve fairness of the LLM.
Critical weaknesses
  1. Numbers in the abstract do not match the results in main content. For example, abstract reports label preferences in 735 of 1,060 conditions (structural 190/363, ordinal 248/315, semantic 297/382) and first-option effects in 265 of 716 comparisons. But table 4, Table 5 and Section 3.3 report 743/1,053 (199/363, 245/309, 299/381) and 265/711. The inconsistencies in the results make me highly skeptical of the soundness of the paper. I hope the authors could clarify these inconsistancy.
Questions for authors

See weakness.

Reviewer_iZydOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how labels, answer attribution, and route advice affect LLM decisions. It tests static choices, judgments about answers labeled as one's own or another model's, and grid navigation across 59 endpoints. It reports stable choice preferences, small attribution effects, and much lower goal success under misleading route advice. The results are organized into ten types of bias.

Critical strengths
  1. The study connects static choices with interactive behavior. The large drop in goal success under misleading advice is useful evidence that following advice and solving the task are different objectives (Figures 5–6).
  2. The attribution tests reuse answer content. Changing its stated source while holding the answer fixed is a useful way to study source effects without changing answer quality.
Critical weaknesses
  1. The design supports narrower claims than the broad framing. Many endpoints are aliases or related snapshots. Attribution tests use only six base questions, and navigation uses two maps without a no-advice baseline. These choices limit task-level generalization and do not isolate a general tendency to follow bad advice from limited information and exploration effects (Appendices A–B).
  2. Poor representativeness of data for claims. Compared to claims in the paper, the author's self-constructed datasets are limited in diversity of base tasks, too simplified in evaluation settings, and lack variation in prompts and labels.
Questions for authors
  1. Clarification: How do the main averages change after grouping aliases and related snapshots, and after excluding repaired outputs on a fixed endpoint set?
  2. Clarification and scope: Are no-advice navigation results available? If not, what conclusion can be supported about the effect of misleading advice relative to independent exploration?
1 comment

I have read the author's response. The new results partly solve my concerns. However, these new results are still insufficient to solve Weaknesses 1 and 2. Therefore, I maintain my scores.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.