acceptodds
Under review as a conference paper at ICLR 2027

Errors and path-dependencies in LLM descriptions of the world: An analysis via maps and routes

Abstract

When a language model describes a piece of the world in prose — a scene, a budget, a walk across a city — it produces a stream of small factual commitments: how many people were in the group, how much is left after the rent is paid, which side of the street the museum is on. We study two surprisingly robust properties of these commitments that are easy to overlook. The first is that models make discrete factual errors at a non-trivial rate even when the same model, asked directly, can identify the error. The second is that the content a model chooses to report is path-dependent: it changes systematically with the order in which material is presented or traversed, in ways that a faithful description of the world should be invariant to. We develop these two themes in a domain that forms the basis for widely-used applications, and where ground truth is unusually clear-cut: descriptions of walking routes through real cities. Using a benchmark of 6032 pedestrian corridors in 35 cities, we evaluate the discrete claims made in the descriptions against map data for ten models across families and sizes, and we find initial error rates on landmark claims between 36% and 72%, with consistency across models in which types of errors are the most frequent. Claims about the existence of landmarks are almost always right; models rarely invent a place outright, but they do misplace it. They name real landmarks from the right part of the city but from outside the stretch being walked, and side-of-street claims from the perspective of the walker are essentially random for smaller models. The failure is one of precision rather than recall, and increasing model capability does not eliminate it: the strongest model we evaluate still gets more than a third of its landmark claims wrong. We then study iterative critique-and-revision as a remedy. We formalize this with a three-parameter model: an initial error rate , a per-revision repair probability , and a per-revision corruption probability . The interaction of and produces an equilibrium minimum error rate that revision cannot improve on. Models differ here by degree rather than by kind: rises with capability from about , where revision merely churns errors, to about , lowering the floor from roughly to but never near zero. A city's difficulty is well predicted by two axes that we can measure independently: how much text about the city exists on the Web, and how regular the city’s street network is. For path-dependence, we show that describing the same corridor in reverse leaves average accuracy unchanged but changes the description itself: the two directions name overlapping sets of landmarks with Jaccard similarity only 0.12–0.35, and for landmarks named in both directions the stated side of the street fails to invert between 30% and 52% of the time. A substantial part of this divergence can be seen even in synthetic cities where every relevant fact is supplied in the prompt and no world knowledge is required, which lets us separate a knowledge-asymmetry component from a pure position narration component. We find evidence for this kind of path-dependence even in stylized narrations in other domains, where for example reordering the categories in a household budget changes how much money the model says the family should be saving. We conclude by discussing what these regularities suggest about how such models represent and deploy knowledge of the world, and what they imply for systems that put model-generated descriptions in front of users.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, maps, navigation, factual errors, path-dependence

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ykGmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies factual errors and path-dependence in LLM-generated long-form descriptions of the world. Using a city-walk benchmark of pedestrian corridors, it shows that models often make incorrect landmark-level factual commitments during free-form narration even when they can often identify the same errors when later asked to audit them. The paper’s main contributions are a large-scale claim-level benchmark with a grounded evaluation pipeline and a three-parameter model of critique-and-revise loops, (p,q,r), used to analyze repair, corruption, and the resulting error floor under repeated revision. Validation relies on external map data, geometric relationships, a grounded automated evaluator with evaluator checks and a human audit, as well as forward/reverse traversal consistency analyses. These experiments provide substantial evidence for the main claims, though in my view some central revision-dynamics conclusions would benefit from stronger robustness analysis across prompting variants.

Critical strengths

The paper introduces a valuable city-walk benchmark that supports large-scale, claim-level, map-grounded analysis. Combined with the grounded evaluation pipeline and the breadth of the city and model comparisons, this enables multidimensional analysis of factual errors across models, cities, and revision rounds, making the empirical phenomena statistically convincing. The paper also provides a clear analytic framework for studying revision dynamics. In particular, the three-parameter model of critique-and-revise loops, (p,q,r), offers an interpretable way to separate repair from corruption and to explain why repeated self-revision exhibits a nonzero error floor. More broadly, the paper meaningfully distinguishes between possessing geographic knowledge and maintaining spatial consistency during long-form narration.

Critical weaknesses

The generation-side procedure has limited methodological novelty. The self-critique-and-revise loop is close to an existing critique-and-revise paradigm, and the paper’s main novelty seems to lie more in the benchmark construction, grounded claim-level evaluation pipeline, and the (p,q,r) analysis than in the revision mechanism itself. This matters for my overall recommendation because a substantial part of the paper’s headline empirical story is organized around this loop, yet the paper does not sufficiently establish what is methodologically new beyond applying a fairly standard paradigm in this domain. The main quantitative conclusions about revision dynamics and the estimated error-rate floor are established under a fairly specific prompt family, especially a full-rewrite revision setting. Different prompt formulations or rewrite instructions could plausibly change both corruption and repair behavior, and full rewrites may themselves introduce additional stochasticity by corrupting previously correct claims. Therefore, I would like to see stronger stability evidence across prompt variants, revision formulations (e.g., full rewrite versus more local editing), decoding settings, or repeated runs. Without such robustness analysis, it remains unclear how much of the reported floor is a general phenomenon of long-form spatial narration and how much is specific to the chosen prompting protocol.

Questions for authors

How sensitive are the fitted revision dynamics and the resulting error-rate floor to the prompting protocol? In particular, if you vary the critique/revision prompt, compare full-rewrite revision with more local editing, or rerun the same routes under multiple prompt variants or decoding settings, do the key qualitative conclusions remain stable—especially the presence of a nonzero corruption rate q and a persistent error floor? This would materially affect my assessment, because it would clarify whether the reported floor is a robust property of long-form spatial narration or is partly induced by the specific critique-and-rewrite setup used in the paper.

2 comments
AuthorsOpenReview ↗

Author Comment

We thank you for the review. We agree with you that having a model revise its own output is an existing family of frameworks, and that our contributions (as noted in the paper) are in the construction of the task and the evaluation, the empirical findings, and the theoretical analysis of the revision framework, which is new relative to prior work. This new analysis reveals things that were not a priori known and are fundamental, including the fact that deletion is the primary source of improvements in these revisions, and that performance on a given city can be predicted from properties of the city. This analysis also provides an optimal algorithm for scheduling revisions across a family of multiple revisers; this formal problem specification and the resulting algorithm are both new.

The evidence in the paper suggests that the phenomena are robust, in a number of respects. First, the prompt for the revision explicitly asks the model to maintain correct landmarks; we note that in the standard paradigm for generation, we can’t replace a token in situ; rather, the generative process must take into account all context. So alternative approaches to this will still have the same constraint of token generation. Second, we did in fact try multiple prompting strategies; some of these are in the paper, where for example Appendix D shows that when we add the true heading after each turn, it does not significantly change side-of-street errors. Others were preliminary runs that are not reported in the current version: in an earlier run with a different prompt, on all 6034 routes, every one of the models again had corruption above zero and a floor in the same range. We can show this variation across prompts in a revision.

If these points have addressed your concerns in part, we would appreciate it if you considered raising your score.

Thank you for the additional clarifications. I appreciate the discussion of prompt variation and the theoretical analysis. However, my concern about novelty remains. I continue to view the primary contribution as the benchmark, empirical findings, and accompanying analysis, rather than a substantially new revision framework. I will therefore keep my original score unchanged.

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

The paper provides a benchmark of 6,032 pedestrian corridors in 35 cities and evaluated 11 closed-book models on them (model are not allowed for tool-usage e.g. browsing/google-map). Each model produced four self-revised drafts, and every named landmark was checked against map data on four binary legs (existence, location, visibility, side of street), yielding 2.5M evaluated claims.

On this benchmark, the paper probes whether LLMs can produce long-form factual descriptions of the physical world and identify two properties of LLM-generated text about the physical world.

  1. accuracy and revision rate: first-draft claims are wrong 41–79% of the time, even when the same model, asked directly, can identify the error. The paper proposed a three-parameter revision model (initial error p, repair r, corruption q) that implies a stationary error rate e* = q/(q+r) of 28–50% that iterative self-critique does not cross, because detection is easy (models flag about four in five of their own errors) while repair is the bottleneck.

  2. path-dependent: it changes with the order in which the environment is traversed, in ways a faithful description should be invariant to. A reversal experiment shows the descriptions are path-dependent — walking the same corridor backwards leaves accuracy unchanged (53.8% vs. 54.3%) but changes what is said, with roughly two-fifths of landmarks named in one direction only and left/right failing to invert.

This paper suggests that models are good at naming real places along a route and putting them in order, but still get relative position wrong: which side a place is on, how left and right swap when the route is reversed, and whether a landmark is actually visible from a given spot.

Critical strengths
  1. The city-walk task focused on in this paper is motivating. They've found that models have numerous deficiencies in their cognitive abilities regarding maps and routes. Accurately identifying these model flaws and improving upon them will be helpful for the fields of navigation and spatial understanding tasks. Therefore, I believe this paper has chosen an excellent starting point and analytical dimension.
  2. This paper explores having the model perform multiple rounds of self-revision to obtain a relatively stable evaluation of its error rate, which has a certain degree of innovation.
Critical weaknesses
  1. On positioning the paper's contributions: This paper states that its first contribution is providing a benchmark featuring over 6,000 walking routes across 35 cities. However, The construction and evaluation of this benchmark might not be very rigorous.

    • the test prompt only asks the model "what are some landmarks I would see along the way, and what side of the street would they be on?". Different models might output varying numbers of landmarks, and the same landmark might have aliases. The paper doesn't seem to detail how these situations should be handled.
    • They have a total of 2.5M evaluated claims, and it uses Gemini 3.8 Flash Lite for extraction. I’m not sure about the confidence level and cost of this evaluation, and I’m wondering if there’s a way to perform a more accurate evaluation with a lower budget?
  2. The readability of this paper needs improvements. The paper contains a lot of weird words without clear definition, which look like they were written by AI (e.g. prose, recur, commitment, salient, attestation). Also, in the introduction, the very first example comes from the appendix, and they don't have much connection to the rest of the text. Core contributions were not listed point by point, but instead written as two paragraphs. This makes it difficult during review to clearly identify the core problem the paper aims to solve and how its methods contribute beyond previous work.

Questions for authors
  1. See critical weaknesses 1.
  2. I’m not entirely clear on what the core message that the paper is trying to convey. If you could list three core contributions/findings that are more human-readable, what would you choose? And what experimental results support them?
2 comments
AuthorsOpenReview ↗

Author Comment

Thank you for your review. We believe that the soundness concerns that you raise are in fact handled by procedures that are described in the paper.

In particular, in W1 you asked about varying numbers of landmarks and the possibility of aliases. In our analysis, accuracy is measured per claim, and so it is normalized to account for the number of landmarks each model names. Aliases are resolved to canonical Google Maps entities (with 98.5% of first-draft claims resolved, as in Appendix B.1–B.2, so this is not a significant source of error). Similarly, claims are matched across drafts and directions by map identifier, falling back to the normalized canonical name (as in Appendix C.3, E.3).

In W1 you also asked about the confidence level and cost. The bootstrap 95% half-widths in Table 1 are at most plus/minus 1.1 points. We evaluated what happens when we rescore approximately 160,000 claims with another evaluator checkpoint, and this changes accuracy by only 0.009%. Moreover, a human audit agrees with the evaluator on 95% of claims (in Appendix B.3). Regarding the cost, the extractor uses a small model that runs once per draft and is cheap, and 3 out of 4 dimensions (existence, distance, and side) are computed from map geometry. Dropping the one judged dimension (visibility) leaves the ranking of all 11 models unchanged (Appendix B.3). So this is a low-budget pipeline, and VLM is only called on visibility and ambiguous cases with marginal tradeoff for accuracy. More generally, the choice of model will generally face a trade-off between cost and performance, but in our case we are using solutions that maintain reasonable cost for the task being performed.

In Q2, you asked about simple statements of three key findings. We would point to: (i) The errors made by the models on maps and routes are about relative location and position rather than being fabrications (since more than 97% of the locations do exist, in 10 of the 11 models). See Table 1. Moreover, the output of even the strongest models is heavily path-dependent, and shows different behavior depending on the order in which the location facts are encountered. See Sections 5.4 and 5.5. (ii) Self-revision can be analyzed, and this analysis gives both efficient algorithms for sequencing revisions and a floor of achievable accuracy. See Appendix C and F. (iii) Performance across different cities can be predicted from properties of the city. See Section 4 and D. We find that this per-city variation is governed by two fundamental dimensions: the structure of the street map of the city, and the volume of data available about the city. These two dimensions can vary separately (e.g. a city might have a complicated street map and a lot of data about it online, or a structurally simple street map and relatively little online data), and they provide a way of understanding how the model’s performance differs significantly across cities. This tells us what kinds of cities are better or worse for geospatial guidance from models.

These points are previewed in the introduction of the submission and several are noted in the "critical strengths" portion of your review; we can make them more visible in a revision.

If these points have addressed your concerns in part, we would appreciate it if you considered raising your score.

Thank you for clarifying the credibility of the assessment and the core points of this article. However, I haven't seen my core concerns on readability (W2, also reviewer#RGVE's W3) being properly addressed in the rebuttal text or the revised PDF. Since your reply was given on the last day of the discussion, I don't have enough time to compare all your modifications. I read the text of the new version, the introduction part is still difficult to follow, and the "conclusion and discussion" part occupy a whole page. I will remain my original score.

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

The paper studies two recurring problems in LLMs when they describe the world in maps and routes tasks. More specifically, 1) making factual mistakes even when the same model can identify the error, and 2) path dependency where the answer changes based on how the environment is traversed. A benchmark or real working routes from Google Maps is designed for this purpose as well as critique-and-revise evaluation loop. Experiments across eleven proprietary models show that all models exhibit non-trivial error rate when they are asked to produce text about maps and routes . Moreover, while model maintain there accuracy when asked about a route in forward and reverse direction, the set of landmarks that it identifies differ significantly.

Critical strengths
  • The paper studies a task that requires spatial reasoning, a skill that LLMs struggle with.

  • Several recent frontier models are evaluated on the paper, which provide insights relevant to the current landscape of LLMs.

Critical weaknesses
  1. The provided analysis about maps and routes (more specifically naming landmarks along corridors) is very limited in scope. While it can be beneficial to understand how LLMs perform in this domain, this should be a part of more exhaustive evaluation of the underlying problems. Otherwise, it becomes difficult to draw conclusions on either this is a domain-specific problem or a broader issue with LLMs.

  2. Since the paper only introduces a benchmark and evaluation, the primary area of the paper is more suited for dataset and benchmark.

  3. I find the writing style of the paper very odd. It reads more like a blog post with dense paragraphs rather than a Machine Learning paper with a defined taxonomy. Also most of the implementation and technical details are in the appendix (e.g. benchmark construction, evaluation protocol which is extremely long to go through. For an informative ML paper, technical details matter more in the main text, while prolonged narratives can be added in the Appendix for further details.

  4. The evaluation in 2.2 relies on a VLM (Gemini 3.8) for reading the map tiles which makes it dependent on the performance of that model. I can see in the Appendix that there was a human audit on 300 samples, but it is not clear of how much does this number of sample constitutes of total number of samples. Also how were they sampled?

  5. In section 5, the paper studies path dependence where the accuracy is unchanged between forward and reverse direction, but only the set of landmarks is what changes. This makes me wonder on what an idea behavior in that case would be. For instance, how would a humans approach and solve this task. Without a clear objective or oracle, it is difficult to draw conclusions from this analysis.

  6. The paper only evaluates closed source models. For this work to be relevant to a wider research community, some open-source models (e.g. QWen, Gemma) shall be included.

  7. After reading the paper, I am still unsure about the main message of the paper. The paper showed that frontier LLMs make mistakes with the specific task at hand, but it is not clear to me how these findings could inform or guide the development of future LLMs.

Questions for authors
  1. At line 132, what is a closed book model?
1 comment
AuthorsOpenReview ↗

Author Comment

Thank you for your review. Here we’d like to address the points you brought up as weaknesses, and your question.

Regarding W1, we do not believe the analysis of maps and routes is limited in scope, given that (i) it is assessing an application domain used by over 2 billion people every day (increasingly, through LLMs), and (ii) it is a task that naturally allows for the assessment of claims at large scales against clear ground truth: in our formulation, with 35 cities, 11 models from five families, and 2.5 million claims. Additionally, the same effects appear without maps. In the story-writing task, reversing order of budget categories, and a synthetic city with every fact in the prompt shows the same direction effect.

Regarding W2, the paper contains many components beyond a benchmark task; it provides a model of self-revision with underlying theoretical analysis and guarantees (in Appendix C and F), governed by a set of parameters that facilitates several lines of analysis. In particular, we are able to use the parameters of the model to understand which parts of the revision process govern the accuracy. Second, we are able to analyze how the revision dynamics and eventual error rate vary across cities, and we find that this per-city variation is governed by two fundamental dimensions: the structure of the street map of the city, and the volume of data available about the city. These two dimensions can vary separately (e.g. a city might have a complicated street map and a lot of data about it online, or a structurally simple street map and relatively little online data), and they provide a way of understanding how the model’s performance differs significantly across cities. This tells us what kinds of cities are better or worse for geospatial guidance from models.

Regarding W4, the conclusions do not hinge on the VLM's judgment. In assessing errors, the evaluation of existence, location, and side are based on facts precomputed from Google Maps data and geometry (Appendix B.1). The only property that uses this additional judgment is visibility, and if you were to disregard visibility, the ranking of all 11 models remains unchanged The precision of the audit depends on the absolute size of the sample (300), not the relative size of the sample compared to the full population, and the sample size is sufficient to generate the tight confidence interval reported in the paper.

Regarding W5, since the set of facts is independent of the direction being walked, the objective is that the reported information should remain the same as well. This is where we see a clear deviation in practice: the stated side flips (as it should) at most only 70.9% of the time, and is near chance for six of the models.

Regarding W6 and Q1, "closed-book" means that we constrain the model to use no retrieval, browsing, or map tools. It does not mean that the model has closed weights. In fact, we use multiple open-weight models in the paper: Qwen 3.8 and GLM-5.3.

Regarding W7, we believe (as discussed in the introduction) that the paper goes significantly beyond the simple fact that LLMs make errors, and addresses both the structure of these errors and a framework for remediation. In particular, findings of this type include: (i) The errors are about relative location and position rather than being fabrications (since more than 97% of the locations do exist, in 10 of the 11 models). (ii) The output of even the strongest models is heavily path-dependent, and shows different behavior depending on the order in which the location facts are encountered. (iii) Self-revision can be analyzed, and this analysis gives both efficient algorithms for sequencing revisions and a floor of achievable accuracy. (iv) Performance across different cities can be predicted from properties of the city.

If these points have addressed your concerns in part, we would appreciate it if you considered raising your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.