Hostile prompts cause 10% drop in LLM instruction following

// 45d agoRESEARCH PAPER

Hostile prompts cause 10% drop in LLM instruction following

A benchmark study of 14 LLM configurations across Llama, Mistral, and Qwen architectures reveals a consistent "hostility residual" where aggressive user framing degrades instruction-following performance. The effect is most pronounced at the 7-8B scale with a 7.4 percentage point drop, persisting even in 123B parameter models despite scaling defenses.

// ANALYSIS

Hostility acts as a "vibe-based" adversarial attack, proving that LLMs are significantly more sensitive to the emotional register of a prompt than previously quantified.

–Scaling models provides a slight defense but fails to eliminate tone-based performance degradation.
–Instruction tuning can amplify sensitivity to hostile framing, suggesting a trade-off between following instructions and emotional robustness.
–Specific model/quantization combinations exhibit emergent position biases under hostile conditions, indicating structural instability in reasoning under "stress."

// TAGS

llmbenchmarkingifevalprompt engineeringinstruction followingmodel robustnessmachine learning

DISCOVERED

45d ago

2026-04-24

PUBLISHED

45d ago

2026-04-24

RELEVANCE

8/ 10

AUTHOR

Saraozte01

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCEjust now

Agent-Reach is an open-source Python tool that equips AI coding agents with zero-config, API-free search and scraping capabilities across platforms like Twitter, Reddit, YouTube, and XiaoHongShu.

Agent-Reach is an open-source Python toolkit designed to give AI agents (like Claude Code, Cursor, and Windsurf) direct web search and scraping abilities without the need for expensive API fees or complex setups. By serving as an installer and routing scaffold, it configures locally managed CLI tools (such as yt-dlp and platform-specific scrapers) so that agents can fetch real-time web content, social media posts, and video transcripts directly. The system emphasizes privacy by keeping credentials and cookies local, and handles smart routing to call the most appropriate tool for a given platform.

OPEN SOURCE1m ago

Personal AI Infrastructure is an open-source framework by Daniel Miessler that acts as a modular "Life Operating System" to scaffold and coordinate personal AI assistants.

Personal AI Infrastructure (PAI) is a comprehensive scaffolding framework that integrates AI tools, context, and personal preferences into a unified platform. It leverages core components like TELOS (for articulating values and goals), Pulse (a local dashboard for monitoring state and work), Skills (functional capabilities), and Memory (contextual history) to create a highly personalized, local AI agent that adapts and improves its assistance over time.

NEWS22m ago

Anthropic clarifies Claude Mythos pricing

Anthropic corrected a pricing misconception regarding its unreleased Claude Mythos model, clarifying that the cost is $25 per million input tokens and $125 per million output tokens rather than a flat monthly fee of "$25/m". Claude Mythos is a highly capable frontier AI model with advanced coding and cybersecurity capabilities, currently restricted to vetted defensive partners via the Project Glasswing initiative.