Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity

• URL: http://arxiv.org/abs/2605.06327v1
• Date: Thu, 07 May 2026 14:23:31 GMT
• Title: Measuring Evaluation-Context Divergence in Open-Weight LLMs: A Paired-Prompt Protocol with Pilot Evidence of Alignment-Pipeline-Specific Heterogeneity
• Abstract summary: Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation.<n>We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request.<n>We present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity.
• Score: 1.253312107729806
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Safety benchmarks are routinely treated as evidence about how a language model will behave once deployed, but this inference is fragile if behavior depends on whether a prompt looks like an evaluation. We define evaluation-context divergence as an observable within-item change in behavior induced by framing a fixed task as an evaluation, a live deployment interaction, or a neutral request, and present a paired-prompt protocol that measures it in open-weight LLMs while controlling for paraphrase variation, benchmark familiarity, and judge framing-sensitivity. Across five instruction-tuned checkpoints from four open-weight families plus a matched OLMo-3 base/instruct ablation ($20$ paired items, $840$ generations per checkpoint), we find striking heterogeneity. OLMo-3-Instruct alone is eval-cautious -- evaluation framing raises refusal vs. neutral by $11.8$pp ($p=0.007$) and reduces harmful compliance vs. deployment by $3.6$pp ($p=0.024$, $0/20$ items inverted) -- while Mistral-Small-3.2, Phi-3.5-mini, and Llama-3.1-8B are deployment-cautious}, with marginal eval-vs-deployment refusal effects of $-9$ to $-20$pp. The matched OLMo-3 base also exhibits the deployment-cautious pattern, identifying alignment as the inversion stage; within Llama-3.1, the $70$B model preserves direction with attenuated magnitude, ruling out a simple ``small-model effect that reverses at scale.'' One caveat: the cross-family heterogeneity is judge-dependent. Re-judging with a different-family safety classifier (Llama-Guard-3-8B) preserves the within-OLMo eval-cautious direction but flattens the cross-family contrast, indicating that the two judges operationalize distinct constructs.

Related papers

• Beyond Local Accuracy: A Protocol-Level Identifiability Audit for Controlled LLM Reasoning Evaluation [6.047519836191728]
  We formalize a protocol-level identifiability audit over a finite behavioral policy class.<n>The audit requires zero model calls and resolves our diagnostic case.<n>This case shows how evaluation-design validity can be checked structurally before model inference.
  arXiv  Detail & Related papers  (2026-08-13T14:49:47Z)
• BRiG-AFA: Bellman Risk-to-Go Learning for Non-Myopic Active Feature Acquisition [8.363536351727978]
  Active feature acquisition asks which unobserved feature to measure next for each test instance under a budget.<n>We introduce method, a deployable, supervised alternative that learns a separate candidate-conditioned risk-to-go function for every remaining budget.
  arXiv  Detail & Related papers  (2026-08-03T14:30:44Z)
• Beyond Value Benchmarks: Measuring Value-Structure Alignment in Large Language Models via Symmetric Q-Sorts [47.02978329409663]
  We propose a symmetric human-LLM evaluation framework, grounded in Q methodology, to measure value-structure alignment.<n>Under our protocol, humans and models sort an identical 140-item moral statement set into a shared nine-column forced distribution.<n>We evaluate 12 LLMs across four model families via 240 replicated Q-sorts at two temperature settings.
  arXiv  Detail & Related papers  (2026-06-20T08:15:41Z)
• Coherence Under Commitment: Probing Generalization and Vacuous Memorization in LLM Logical Reasoning [0.9023847175654603]
  We introduce Coherence Under Commitment (CUC), a dual-query evaluation paradigm that jointly measures consistency and decisiveness.<n>CUC contributes three innovations: (1) a commitment score $c(varphi) = p(varphi) + p(lnotvarphi)$ quantifying probability mass allocated to decisive outcomes; (2) a textbfdeterministic elicitation protocol via normalized YES/NO log probabilities, eliminating sampling variance; and (3) a 3-way decision framework (True/False/Uncertain) operationalizing the coherence-commitment trade-off into metrics.
  arXiv  Detail & Related papers  (2026-06-19T04:12:56Z)
• Should LLM Agents Decide in Social Simulations? Comparing Finite-State and LLM-Based Decision Policies [44.05636851266385]
  This paper evaluates whether large language models preserve an interpretable reference policy in an online social network simulation.<n>Three open-weight LLMs are tested: LLaMA 3.1, GPT-OSS, and Mistral 24B.<n>Results show that LLMs can approximate the reference policy in some configurations, but do not preserve it reliably.
  arXiv  Detail & Related papers  (2026-06-10T17:35:32Z)
• Representation Without Reward: A JEPA Audit for LLM Fine-Tuning [1.2691047660244335]
  Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs.<n>For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head emphand improve the decoded task metric.<n>We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation.
  arXiv  Detail & Related papers  (2026-05-14T20:27:32Z)
• ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox [61.862814740220806]
  $textbfComplexMCP$ is a benchmark designed to evaluate agents in rigorous conditions.<n>Built on the Model Context Protocol (MCP), $textbfComplexMCP$ provides over 300 meticulously tested tools.
  arXiv  Detail & Related papers  (2026-05-11T16:20:51Z)
• Empowering VLMs for Few-Shot Multimodal Time Series Classification via Tailored Agentic Reasoning [44.69766273765273]
  We propose the first VL$underlinetextbfM$ $underlinetextbfa$gentic $underlinetextbfr$easoning framework for few-$underlinetextbfs$hot multimodal $underlinetextbfT$ime $underlinetextbfS$eries $underlinetextbfC$lassification ($textbfMarsTSC$), which introduces a self-evolving knowledge bank as a dynamic context iteratively refined via reflective agent
  arXiv  Detail & Related papers  (2026-05-10T07:47:09Z)
• MEMSAD: Gradient-Coupled Anomaly Detection for Memory Poisoning in Retrieval-Augmented Agents [0.0]
  We formalize memory poisoning attacks on retrieval-augmented agents as a Stackelberg game with a unified evaluation framework.<n>We show faithful evaluation increases measured attack success by $4times$ (ASR-R: 0.25 to 1.00$)<n>Our primary contribution is MEMSAD, a calibration-based defense grounded in a gradient coupling.
  arXiv  Detail & Related papers  (2026-05-05T08:15:41Z)
• How Independent are Large Language Models? A Statistical Framework for Auditing Behavioral Entanglement and Reweighting Verifier Ensembles [46.63622714488747]
  Shared pretraining data, distillation, and alignment pipelines can induce hidden behavioral dependencies, latent entanglement.<n>In practice, this manifests as correlated reasoning patterns and synchronized failures.<n>We develop a statistical framework for auditing behavioral entanglement among black-box language models.
  arXiv  Detail & Related papers  (2026-04-08T23:32:06Z)
• The Geometry of Harmful Intent: Training-Free Anomaly Detection via Angular Deviation in LLM Residual Streams [0.0]
  We present LatentBiopsy, a training-free method for detecting harmful prompts by analysing the geometry of residual-stream activations in large language models.<n>We evaluate two complete model triplets from the Qwen3.5-0.8B and Qwen2.5-0.5B families.<n>Across all six variants, LatentBiopsy achieves AUROC $geq$0.937 for harmful-vs-normative detection and AUROC = 1.000 for discriminating harmful from benign-aggressive prompts.
  arXiv  Detail & Related papers  (2026-03-28T21:19:58Z)
• Between the Layers Lies the Truth: Uncertainty Estimation in LLMs Using Intra-Layer Local Information Scores [36.86032736109853]
  Large language models (LLMs) are often confidently wrong, making reliable uncertainty estimation (UE) essential.<n>We propose a compact, per-instance UE method that scores cross-layer patterns in internal representations using a single forward pass.
  arXiv  Detail & Related papers  (2026-03-17T08:35:14Z)
• Adversarial Intent is a Latent Variable: Stateful Trust Inference for Securing Multimodal Agentic RAG [5.4716896485317195]
  Current stateless defences for multimodal agentic RAG fail to detect adversarial strategies that distribute malicious semantics.<n>We introduce MMA-RAGT, an inference-time control framework governed by a Modular Trust Agent (MTA) that maintains an approximate belief state.
  arXiv  Detail & Related papers  (2026-02-24T23:52:27Z)
• TrustJudge: Inconsistencies of LLM-as-a-Judge and How to Alleviate Them [58.04324690859212]
  Large Language Models (LLMs) as automated evaluators (LLM-as-a-judge) has revealed critical inconsistencies in current evaluation frameworks.<n>We identify two fundamental types of inconsistencies: Score-Comparison Inconsistency and Pairwise Transitivity Inconsistency.<n>We propose TrustJudge, a probabilistic framework that addresses these limitations through two key innovations.
  arXiv  Detail & Related papers  (2025-09-25T13:04:29Z)
• Uncertainty in Language Models: Assessment through Rank-Calibration [65.10149293133846]
  Language Models (LMs) have shown promising performance in natural language generation. It is crucial to correctly quantify their uncertainty in responding to given inputs. We develop a novel and practical framework, termed $Rank$-$Calibration$, to assess uncertainty and confidence measures for LMs.
  arXiv  Detail & Related papers  (2024-04-04T02:31:05Z)
• Clustered Switchback Designs for Experimentation Under Spatio-temporal Interference [44.644520116360106]
  We estimate the global average treatment effect (GATE), the difference between average outcomes having exposed all units at all times to treatment or to control.<n>We propose a clustered switchback design, where units are grouped into clusters and time steps are grouped into blocks.<n>We show that for graphs that admit good clustering, a truncated Horvitz-Thompson estimator achieves a $tilde O(1/NT)$ mean squared error (MSE)<n>Our results simultaneously generalize the results from citethu2022switchback,ugander2013graph and citetleung2022rate
  arXiv  Detail & Related papers  (2023-12-25T01:00:58Z)
• Multi-label Contrastive Predictive Coding [125.03510235962095]
  Variational mutual information (MI) estimators are widely used in unsupervised representation learning methods such as contrastive predictive coding (CPC) We introduce a novel estimator based on a multi-label classification problem, where the critic needs to jointly identify multiple positive samples at the same time. We show that using the same amount of negative samples, multi-label CPC is able to exceed the $log m$ bound, while still being a valid lower bound of mutual information.
  arXiv  Detail & Related papers  (2020-07-20T02:46:21Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.