When Context Misleads: In-context Learning with Jurisdiction in Large Language Models
Pei-lin Li, Qingle Liu, Junyang Feng, Siyu Li, Sunqi Fan, Xin-Sheng Chen, Shuojin Yang
cs.CL · cs.AI
Quick verdict
Benchmark + J-ICL post-training claim ICL and fake-context resistance improve together; well-audited but single-seed, no CIs, thin Qwen3-4B gains.
Summary
This paper introduces context-authority discrimination: whether an LLM should let few-shot demonstrations govern its answer. The authors release FakeContextBench, 3,500 instances from 700 theory families across seven domains, where coherent but pseudoscientific demonstration rules conflict with reality-grounded answers, scored by a deterministic judge validated against humans (90% agreement, kappa 0.807). Evaluations show scaling does not confer resistance: Qwen3 base models gain 27.40 points of ICLEval from 0.6B to 14B while Reality Rate falls 3.46 points, and ICL post-training can hurt — MetaICL and Symbol Tuning reduce reality accuracy by up to 14.95 points (Symbol Tuning on Qwen3-8B). The proposed J-ICL post-training jointly trains applicable (J+) and inapplicable (J−) episodes with a query-only loss and no explicit authority labels. Across four backbones it averages +5.84 ICLEval and +9.20 reality-accuracy points over base, and +18.09 Reality Rate over the MetaICL/Symbol-Tuning average. Gains survive paraphrase and semantic-cue shortcut audits (96.77% prediction invariance). Limitations: single-seed training without confidence intervals, no Reality-Rate gain on Qwen3-4B (27.43 vs 27.66), hyperparameters chosen to favor ICLEval over robustness, and activation-patching effects within standard error.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 0 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Model scaling and large-scale pretraining do not reliably produce context-authority discrimination: Qwen3 base models gain 27.40 ICLEval points from 0.6B to 14B while Reality Rate falls 3.46 points.
- Prevalent ICL fine-tuning methods increase susceptibility to misleading demonstrations, with MetaICL and Symbol Tuning reducing reality accuracy by up to 14.95 percentage points (Symbol Tuning on Qwen3-8B).
- J-ICL post-training jointly improves conventional ICL (ICLEval +5.84 average) and reality-grounded accuracy (+9.20 average) over base models across four backbones, and raises Reality Rate by 18.09 points over the MetaICL/Symbol-Tuning average.
- J-ICL's gains persist under paraphrased statements and an adversarial semantic-cue test (96.77% paired-prediction invariance), arguing against template memorization or a negative-cue-to-Reality shortcut.
- Misleading demonstrations act through a partially concentrated late-layer (L33-35) Context-favoring pathway in Qwen3-4B, supported by activation replacement on 209 held-out families.
Strengths
Novel, cleanly formalized problem
Context authority is cleanly distinguished from faithfulness/knowledge-conflict work via dual registered answers yR and yP (§3.1); the benchmark isolates rule-application errors from factual ignorance, and the 'motion force' case (App. A) makes the failure mode concrete.
Unusually well-audited deterministic scorer
Rule-based judge with a documented decision-region cascade and full decision-path logging (App. B.1), validated against human annotators on 100 mixed cases: 90.0% agreement, kappa 0.807 (92.8%, kappa 0.856 on R/C cases).
Shortcut resistance is tested, not asserted
Honest, internally consistent reporting
Ablation isolates the pairing mechanism
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets a blind spot between two literatures: context-faithfulness and knowledge-conflict work assumes external context should win (Zhou et al. 2023; Zheng et al. 2023), while this paper asks when demonstrations should *not* govern an answer. It formalizes 'context authority' via dual registered answers (yR reality-grounded, yP context-induced; §3.1), releases FakeContextBench (700 theory families x 5 instances = 3,500 items across seven domains), and proposes J-ICL post-training. Central claims: (1) scale alone does not yield authority discrimination — Qwen3 base ICLEval rises 27.40 points from 0.6B to 14B while Reality Rate falls 3.46 (Fig. 2a); (2) ICL post-training can backfire — reality accuracy drops by up to 14.95 points (Symbol Tuning on Qwen3-8B, Table 3); (3) J-ICL averages +5.84 ICLEval and +9.20 reality-accuracy points over base across four backbones, and +18.09 Reality Rate over the MetaICL/Symbol-Tuning average. I recomputed all three aggregates from Table 3; they are arithmetically consistent, which is not a given in this genre.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 0 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FakeContextBench?
A 3,500-instance benchmark (700 theory families across seven domains) testing whether LLMs resist following coherent but pseudoscientific rules demonstrated via few-shot examples. Each item pre-registers a reality-grounded answer and a context-induced answer, and a deterministic rule-based judge — validated against human annotators (90% agreement, Cohen's kappa 0.807 on 100 sampled cases) — labels responses as Reality, Context, or Other.
What is J-ICL (Jurisdiction In-Context Learning)?
Jurisdiction In-Context Learning is a post-training method that jointly trains on applicable (J+) and inapplicable (J−) demonstration episodes with a loss applied only to the final query answer, teaching models when contextual rules should govern without explicit authority labels. Across four model backbones it improves ICLEval by about 5.84 points and reality accuracy by about 9.20 points on average over base models.
Does in-context fine-tuning make LLMs more vulnerable to misinformation?
On this paper's benchmark, yes for the tested methods: MetaICL and Symbol Tuning increased susceptibility to misleading few-shot demonstrations, with reality accuracy dropping by up to 14.95 percentage points (Symbol Tuning on Qwen3-8B), and MetaICL reduced Reality Rate relative to the base model on all four backbones tested. The proposed J-ICL method is designed to reverse this effect.
Does making models bigger fix following false context?
Not according to this paper: Qwen3 base models from 0.6B to 14B improved conventional ICL (ICLEval +27.40 points) but their Reality Rate slightly declined (−3.46 points), and among frontier models only GPT-5.6 Sol scored above 50% Reality Rate (55.08%); the others ranged from 13.56% (Gemini 3.5 Flash) to 44.78% (Kimi K3).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 2-3 are available only as OCR text; plotted values were transcribed from text and could contain extraction artifacts.
- Training code release is not explicitly stated; only the benchmark repository (github.com/peilin717/FakeContext-Bench) is cited.
- Frontier-model rows (Table 1) reference 2026 commercial APIs whose evaluation configurations cannot be verified from the provided text.
- Appendices end at J; any further supplementary material (per-domain results, licensing) was not present in the extraction.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.