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

• URL: http://arxiv.org/abs/2608.17247v1
• Date: Tue, 18 Aug 2026 01:04:45 GMT
• Title: Explicit State Elicitation Is Not Enough: A Controlled Audit of Memory-Policy Classification
• Abstract summary: We use this setting to develop an empirical audit protocol for structured intermediate outputs.<n>A 480-example synthetic development set initially suggested large gains from a state-structured prompt bundle.<n>On this set, exposing the four state definitions improves accuracy, but an isolated explicit state-output field does not significantly improve policy accuracy.
• Score: 11.991995412984503
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• A distribution-free certification framework for trustworthy crash-severity prediction [1.089614199781423]
  We develop a certification layer that wraps any severity model unmodified.<n>On 5.2 million Texas records across seven base models spanning four decades, the layer attaches identical validity.<n>The framework is released as an open-source package with theorem-level tests.
  arXiv  Detail & Related papers  (2026-09-10T14:20:08Z)
• CVE-SAI: Counterfactual Visual Evidence-Guided Selective Attribute Indexing for Risk-Controlled E-commerce Search [12.763061404384601]
  Multimodal product models can complete missing e-commerce attributes.<n>Current methods still optimize attribute-answer accuracy, conflate transient prediction with persistent index admission.<n>We address these gaps with Counterfactual Visual Evidence-Guided Selective Attribute Distortioning.
  arXiv  Detail & Related papers  (2026-08-25T18:10:36Z)
• Discourse-Role Labels as Presentation-Time Variables for Context Use in Language Models [3.376024929309146]
  We study the effect of Context-augmented language model systems on reader-model behavior.<n>We introduce a paired fixed-content probe over 500 MMLU-Pro items.<n>Each item receives the same misleading answer-bearing assertion under different discourse-role labels.<n>Across GPT-5.5, DeepSeek V4 Pro, Llama-3-8B-Instruct, and Qwen2.5-7B-Instruct, Misleading Adoption Rate shifts by 56-84 percentage points.
  arXiv  Detail & Related papers  (2026-06-02T18:12:57Z)
• Conformal Certification of Reasoning Trace Prefixes [30.432255322700126]
  We introduce CROP (Conformal Reasoning Output Prefixes), a verifier-agnostic calibration procedure for clean- annotated certification.<n>We demonstrate that CROP over- and under-withholding, improving downstream repair accuracy by preserving valid intermediate reasoning.
  arXiv  Detail & Related papers  (2026-05-28T15:31:42Z)
• Verifier-Bound Communication for LLM Agents: Certified Bounds on Covert Signaling [0.0]
  Colluding language-model agents can hide coordination in messages that remain policy-compliant at the surface level.<n>We present CLBC, a protocol where generation and admission are separated.<n>We show how this protocol yields an upper bound on transcript leakage in terms of latent leakage plus explicit residual channels.
  arXiv  Detail & Related papers  (2026-02-27T23:42:37Z)
• Conditional Coverage Diagnostics for Conformal Prediction [47.93989136542648]
  We show that conditional coverage estimation can be a classification problem.<n>We call the resulting family of metrics excess risk of the target coverage (ERT)<n>We release an open-source package for ERT as well as previous conditional coverage metrics.
  arXiv  Detail & Related papers  (2025-12-12T18:47:39Z)
• COIN: Uncertainty-Guarding Selective Question Answering for Foundation Models with Provable Risk Guarantees [51.5976496056012]
  COIN is an uncertainty-guarding selection framework that calibrates statistically valid thresholds to filter a single generated answer per question.<n>COIN estimates the empirical error rate on a calibration set and applies confidence interval methods to establish a high-probability upper bound on the true error rate.<n>We demonstrate COIN's robustness in risk control, strong test-time power in retaining admissible answers, and predictive efficiency under limited calibration data.
  arXiv  Detail & Related papers  (2025-06-25T07:04:49Z)
• When Does Confidence-Based Cascade Deferral Suffice? [69.28314307469381]
  Cascades are a classical strategy to enable inference cost to vary adaptively across samples. A deferral rule determines whether to invoke the next classifier in the sequence, or to terminate prediction. Despite being oblivious to the structure of the cascade, confidence-based deferral often works remarkably well in practice.
  arXiv  Detail & Related papers  (2023-07-06T04:13:57Z)
• Conservative Prediction via Data-Driven Confidence Minimization [70.93946578046003]
  In safety-critical applications of machine learning, it is often desirable for a model to be conservative. We propose the Data-Driven Confidence Minimization framework, which minimizes confidence on an uncertainty dataset.
  arXiv  Detail & Related papers  (2023-06-08T07:05:36Z)
• Delving into Probabilistic Uncertainty for Unsupervised Domain Adaptive Person Re-Identification [54.174146346387204]
  We propose an approach named probabilistic uncertainty guided progressive label refinery (P$2$LR) for domain adaptive person re-identification. A quantitative criterion is established to measure the uncertainty of pseudo labels and facilitate the network training. Our method outperforms the baseline by 6.5% mAP on the Duke2Market task, while surpassing the state-of-the-art method by 2.5% mAP on the Market2MSMT task.
  arXiv  Detail & Related papers  (2021-12-28T07:40:12Z)
• Distribution-free uncertainty quantification for classification under label shift [105.27463615756733]
  We focus on uncertainty quantification (UQ) for classification problems via two avenues. We first argue that label shift hurts UQ, by showing degradation in coverage and calibration. We examine these techniques theoretically in a distribution-free framework and demonstrate their excellent practical performance.
  arXiv  Detail & Related papers  (2021-03-04T20:51:03Z)

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.