# Learning When to Trust via Selective Context Preference Optimization

Source: [arXiv](https://arxiv.org/abs/2608.06377v1)  
Feed7 permalink: https://feed7.dev/p/2608-06377v1-0rvbpra  
Published: 2026-08-06T17:59:58.000Z  
Trust: Needs Review (needs_review)

## Why Included

MIST tests whether models use good context while resisting bad context, exposing agents that appear robust only because they ignore external evidence altogether.

## Source Summary

MIST renders each reasoning item under **four matched conditions**: clean, misleading, correct-context, and irrelevant-context. Its **SC2W** metric counts cases where misleading context flips an otherwise correct answer to wrong.

## Practical Implication

Evaluate retrieval-augmented agents for selective trust, not only prompt-injection resistance. The proposed **SCOPE** method trains on matched preference pairs balanced across all four conditions so resistance does not come from ignoring useful context.

## Agent-Ready Context

MIST renders each reasoning item under **four matched conditions**: clean, misleading, correct-context, and irrelevant-context. Its **SC2W** metric counts cases where misleading context flips an otherwise correct answer to wrong.

Evaluate retrieval-augmented agents for selective trust, not only prompt-injection resistance. The proposed **SCOPE** method trains on matched preference pairs balanced across all four conditions so resistance does not come from ignoring useful context.

The abstract reports reduced susceptibility on popular open models while preserving other-condition accuracy, but provides no numerical effect sizes here. Broader generalization beyond the benchmark's reasoning items remains an open question from the supplied material.

## Connected Context

Feed7 judgment across 390 accumulated Signals:

This turns context robustness into a selective-trust problem: an agent must resist misleading material without becoming insensitive to correct evidence. The matched four-condition design and per-item SC2W flips sharpen evaluation beyond aggregate accuracy, while the supplied evidence does not establish effect size or transfer beyond the benchmark.

- [The Illusion of Robustness: Aggregate Accuracy Hides Prediction Flips under Task-Irrelevant Context](https://feed7.dev/p/2607-12963v1-1oc0qmr) — MIST operationalizes the earlier warning about hidden per-item flips and extends it by separating misleading, irrelevant, correct, and clean context rather than testing irrelevant noise alone.
- [Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs](https://feed7.dev/p/2607-12985v1-1yo2sej) — Both separate resistance from useful updating; SCOPE trains this balance with matched preferences, while Resist and Update reports a training-free control whose deployable form lost resistance.
- [MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical Sycophancy in LLMs](https://feed7.dev/p/2608-02520v1-16negmk) — MedPRESS tests whether judgment survives escalating interpersonal pressure, complementing MIST’s test of whether judgment survives misleading retrieved context.

## Context Map

- Layer: benchmark
- Domains: research
- Topics: agent-evals, agent-reliability, context-engineering

## Uncertainty

- The abstract reports reduced susceptibility on popular open models while preserving other-condition accuracy, but provides no numerical effect sizes here. Broader generalization beyond the benchmark's reasoning items remains an open question from the supplied material.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.