CS Paper Radar

Right Number, Wrong State? Measuring Cross-Jurisdiction Substitution in LLM Recall of State Policy

Jiayu Feng

cs.CL ·submitted 2026-10-07 05:12 UTC · v2
unusual

Abstract

When an LLM answers a state-specific policy question wrongly, it may be hallucinating, or it may be returning a real value that holds in another state. We test this with a minimal-set design: the question wording is fixed and only the jurisdiction varies, across the 50 U.S. states and the District of Columbia (51 jurisdictions) and three exactly defined Medicaid income-eligibility quantities. Gold values come from an official data book and agree with an independent source in 101 of 102 checked cells. Under a pre-registered protocol, Claude Sonnet 5.5 and GPT-5.6 Sol reproducibly give another state's current value, identical across two independent repeats, for 10 and 25 of 153 items. The count depends on the reference set: crediting any wrong answer that equals another state's value yields 3-5x more reproducible substitutions than checking every number in the asked state's own records, because many apparent cross-state answers are the asked state's own values under another convention or from an earlier year. Claims about cross-jurisdiction error need a complete same-state reference set. We will release the protocol, gold table, and all model outputs.

Comment
6 pages, 3 figures, 1 table

Closest papers

Nearest neighbours by meaning across the whole index.

GT-FENG: General Theory of Factual Evaluation and Normative Grounding in Language Models

cs.CL ·2026-11-13 ·Jiayu Feng
code/data similarity 0.57
Abstract · 7 pages, 4 tables. Code: https://github.com/VonFlair/gtfeng

Whether a language model's answer is correct often depends on context: the same value can be right under one jurisdiction, date, or legal definition and wrong under another. We introduce GT-FENG (General Theory of Factual Evaluation and Normative Grounding), the first general framework to unify contextual/temporal correctness and rule-based error attribution across law, policy, and other rule-governed domains. GT-FENG checks answers against the applicable rules and identifies documented rules that match incorrect answers. It shows that when multiple rules give the same value, the answer alone cannot tell which rule the model used. In Medicaid income-limit answers from two models, documenting the asked state's own records reduces the number of wrong answers that match only other states' values by 68-78%, and more than half of stable wrong answers then match two or more documented sources. GT-FENG provides a foundation for evaluating language models in rule-governed domains and for future benchmarks and evaluation toolkits. We release GT-FENG as an open-source Python package, gtfeng.

0.28

Follow the Policy, Fail the Benchmark: Auditing and Revising Policy-Based Decision Benchmarks

cs.CL cs.SE ·2026-11-04 ·Yanru Xiao, Yifei Cai
benchmarkcode/data similarity 0.53
Abstract · 28 pages. Code and revised data: https://github.com/SugarRuy/BAR-Policy-Decision

Policy-based decision benchmarks evaluate whether systems apply written rules to financial and operational cases. Yet a rule-following response can be marked wrong when the benchmark misstates the required decision. We audit nine public policy-based decision benchmarks through source tracing, program execution, and controlled checks. We find unsupported targets, reference errors that survive replay or propagate as solved steps, and evaluators that reject acceptable answers or credit incomplete decisions. To measure the scoring impact, we rescore unchanged model responses. On the 96 LegalBench SARA numeric cases, correcting the scorer raises acceptance by more than 80 percentage points for each of three models, without changing references or responses. In RuleArena Tax, restoring a required zero floor changes 27 of 300 references and narrows an observed model gap. A controlled federal-exemption task with explicit ownership and eligibility assumptions also shows that errors in supplied intermediate answers can affect final decisions. We present BAR (Benchmark Audit and Revision), a versioned collection of audit evidence and revisions for reevaluation.

0.44

Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification

cs.AI ·2026-08-18 ·Yihang Chen, Pin Qian, Su Wang, Chong Peng, Huan Xu +2 more
benchmark similarity 0.48
Abstract · 34 pages, 1 figure

Personalized agents must decide whether retrieved user memory should be used, ignored, updated, or queried before it affects a current task. We use this setting to develop an empirical audit protocol for structured intermediate outputs: first audit dataset shortcuts, then isolate bundled prompt changes, check whether intermediate labels are answer-associated, test decomposed semantic evidence, and audit provider-level execution failures. A 480-example synthetic development set initially suggested large gains from a state-structured prompt bundle, but TF-IDF diagnostics showed lexical separability and no positive standalone Ignore cases. We therefore construct a frozen 160-example controlled counterfactual set with 40 matched four-way families and rule-derived reference policies. On this set, exposing the four state definitions improves accuracy, but an isolated explicit state-output field does not significantly improve policy accuracy for Llama-3.3-70B and gives only a marginal, non-significant gain for GPT-OSS-120B. Supplying benchmark-associated state labels shifts policy predictions, but because those labels deterministically map to policies, this is a label-conditioning diagnostic rather than evidence of a faithful internal mechanism. Family-level and seed-stability analyses further show that example-level accuracy overstates counterfactual consistency: complete four-way family success is rare. An exploratory follow-up that elicits decomposed semantic evidence also fails to improve routing for the cleanly evaluated endpoint; the corresponding GPT-OSS condition was unavailable because of provider-side request validation. We evaluate policy classification only, not downstream responses, tool actions, or memory-store mutation.

0.54

Jev-as-a-Judge for RAG Claim Verification: Confidence-Gated Cascading Between a Non-Generative Verifier and a Frontier LLM

cs.CL ·2026-09-30 ·Alden Do Rosario, Kirollos Atef
code/data similarity 0.48
Abstract · 31 pages, 6 figures. Code and data: https://github.com/adorosario/jev-rag-claim-verification

Retrieval-augmented generation asks one question at volume: does this evidence support this claim? The answer is one bit, yet most pipelines obtain it from a text generator and pay generation prices for the prose wrapped around it. We compare a non-generative decision model, which returns a typed value from a declared enumeration together with a native probability and no text, against a frontier LLM configured as an engineer would deploy it for binary classification. Both answer the same 495 human-labelled claims from LLM-AggreFact, so every comparison is within-item. We report three things. First, the two systems are statistically indistinguishable on the metric this task is conventionally scored by, a paired difference of -0.56 points, Jev minus Astra, at a 187-fold cost difference. Second, that tie carries almost no information about deployment: the systems disagree on 15.2% of claims, and the cheaper one admits 23.2% of the claims annotators marked unsupported against the frontier model's 14.1%. A component whose job is to catch unsupported claims should be judged on that rate, and the conventional headline hides it. Third, because the errors differ they compose: routing the claims the cheap model is confident about and escalating the rest reaches 74.10 at USD 2.92 per thousand claims against 73.83 at USD 8.93, with the threshold chosen out of fold. But a matched random-escalation control at in-sample thresholds, spending the same budget by coin flip, does as well, so what buys the saving is the escalation share and not the probability that picks what to escalate. Measured against the cheap model it is built from, the cascade costs 61-fold more for 0.83 points, 95% CI [-3.36, 4.86], and it is the more permissive guard of the two. Equal accuracy conceals unequal errors. All predictions, prompts, pricing and code are released.

0.73

norregel: a preregistered benchmark of dated numeric values in Norwegian law

cs.CL ·2026-10-21 ·Eirik Botten Nicolaysen
benchmarkunusual similarity 0.48
Abstract · 24 pages. Preregistrations, code and results: Zenodo concept DOI 10.5281/zenodo.23422302 (versions v1 10.5281/zenodo.23422303, v1.1 10.5281/…

norregel asks language models for numeric values in Norwegian statutes, regulations and parliamentary decisions (NOK amounts, percentages, multiples of the National Insurance basic amount (G) and age limits), each at a stated date. The 916 items were extracted mechanically from Lovdata's open data, and each carries the dated history of its provision, so an answer can be classed as correct, outdated, wrong or uninterpretable. We report a first and a second preregistered run on the same items. In the first run (five models, listed publicly only after local analysis), H1, that outdated answers are more frequent for values that took effect after a model's cutoff, could not be tested in substance, and H2 and H3 were inconclusive; Opus 5.5 and Sonnet 4.6 answered 46.7% and 40.2% of the 880 primary items correctly. The second run was preregistered and published before any of its calls, and eight models were analysed. Its pooled primary test supported H1 under the preregistered decision rule (d = 17.8 percentage points, 95% interval [7.7, 28.9], p = 0.0001). In the preregistered sensitivity analysis S8, which counts as outdated only the most recent value that differs from the key, d was 3.9 [-3.6, 12.7] (p = 0.1754), an interval that contains 0. H1 was supported in every other preregistered sensitivity analysis, but S8 is the only one that gives both groups one target value, so we do not regard the support for H1 as robust. The share of outdated answers equal to that most recent value (H2) was 56.1% [46.6, 65.0], an interval that contains 0.5, and the pooled test of a lower share correct in Nynorsk than in Bokmål (H3) was inconclusive. Several models, among them Haiku 4.5, gave mostly uninterpretable answers. Items and keys are withheld; item IDs, per-answer classes, code and results are on Zenodo.

0.48

Decomposing Wrong-Consensus Agreement in LLM Self-Consistency

cs.CL cs.AI ·2026-08-19 ·Lizhuo Zhang, Mengmeng Tang, Chenfeng Long, Xiaoyong Tang, Xiang Luo
benchmark similarity 0.47
Abstract · v2: Added controlled open-weights replication (five models, ten cells), finite-donor bias calibration, and within-family regime analysis

Agreement among repeated samples of a language model is routinely read as evidence about answer reliability, yet wrong answers can agree just as strongly as right ones. This paper asks what information wrong-consensus agreement actually contains, and answers with a quantitative decomposition. A pluralistic agreement index Gamma, normalized by the reference scale d=(1-p)/(C-1), is split into a mechanical component (agreement delivered by a per-case answer preference alone) and a preference-unexplained residual. The mechanical reference is leak-free: each case's preference and accuracy are estimated from its other runs only. On public GPT-4.1 per-run data, coverage phi (the mechanical/empirical ratio) shows a benchmark-associated direction: 0.81-0.93 on multiple-choice GPQA-Diamond against 0.59-0.78 on open-domain AIME, where a residual of 1.54-2.80 Gamma units survives, more than absorbed by a calibrated run-level preference-heterogeneity reference. A controlled replication under one fixed protocol (four runs per question, K=32 votes) on five open-weights checkpoints (Qwen3.5-9B/122B, Qwen3.8-27B, Gemma4-26B/31B) finds near-complete mechanical coverage in all ten cells (phi approximately 1, with a small overshoot consistent with a quantified finite-donor plug-in bias), robust to a two-run design; the largest cell (qwen3.5-122b, p=0.222) sits inside the GPT-4.1 AIME accuracy range and still saturates (phi=1.041). A cross-system contrast at comparable aggregate accuracy contrasts near-complete mechanical agreement in the open-weights models against a larger preference-unexplained residual in the frontier family. This contrast is confounded with sampling protocol by design. Agreement is graded evidence, not certification. No new voting method is proposed; code and evidence are committed.

0.50

Diagnosing Sampled LLM Reasoning in Formal Geometry: Coverage, Realization, and Validity Evidence

cs.AI ·2026-09-26 ·Xiao Yue, Guangzhi Qu
similarity 0.47
Abstract

Repeated sampling can reveal a correct numerical answer without yielding either a reliable system output or a supported derivation. We present Coverage, Realization, and Validity Evidence (CRV), an evaluation protocol for sampled large language model (LLM) reasoning over formal geometry states. Coverage is answer availability, realization is readout accuracy on the frozen candidate pool, and validity evidence is a label-blinded critic judgment of derivational support rather than a proof certificate. CRV freezes each candidate pool before comparing readouts and analyzes covered failures by correct-answer multiplicity and within-problem discrimination. On HardShift441, a 441-problem set for which a reference solver leaves 406 problems unsolved, a LoRA-adapted Qwen2.5-7B generator obtains 24.2% average single-sample accuracy and 68.9% pass@16, whereas verifier-weighted self-consistency (WSC) reaches 38.0%. Readout accuracy is particularly low when the correct answer occurs only once or twice in the pool. In a separate constructed audit of 195 covered problems, the critic labels 12 correct-answer representatives as supported, 181 as refuted, and two as uncertain. These results show that coverage, realization, and validity evidence from the critic are distinct quantities and should be reported separately.

0.29

Three Ways Classical Test Theory Can Mislead About LLM Judges

cs.LG cs.CLstat.ME ·2026-08-31 ·Louis Yiven Zhu
code/data similarity 0.46
Abstract · 16 pages (7 of main text), 4 figures. v2 adds the gold-computed null for all three statistics and a decision table, corrects the reading of …

Evaluations that use a large language model (LLM) as a judge have begun to borrow reliability statistics from classical test theory and its extensions. We examine three such statistics that need one administration and no gold labels. None of them can isolate the judge, because one judge under one prompt supplies no variance component of its own. Claude Haiku 4.5 judged 210 constructed short answers against ten-element checklists. On the 180 with parsed verdicts, the Kuder-Richardson coefficient (KR-20) came out at 0.5223 on the judge's verdicts and 0.5231 on error-free gold verdicts. In simulation, bank design alone moves KR-20 from 0.01 to 0.68 at the judge's measured 4.72% error rate. The dependability index $Φ(λ)$, a ratio of mean squared distances from the pass mark, sits 0.22 to 0.38 below the judge's accuracy against gold and returns 0.54 to 0.68 on error-free gold verdicts. Livingston-Lewis accuracy treats the rubric elements as a sample, and at a pass mark of five elements it credits error-free gold scores with 0.78, close to the judge's 0.81. A statement about the judge therefore needs gold labels or a varied scorer facet, and a reliability ratio needs the bank's spread beside it. One of the four closest judge-evaluation papers varies the prompt and still reads a reliability below 0.7 as a sign that a model cannot serve as a judge, although that reliability moves with the spread of the samples scored. We derive a decision table and four reporting lines from these two rules.

0.58