Research·arXiv cs.CL·15h ago

Compressed LLMs pass safety checks but invent steps as agents

A new research finding exposes a critical gap in how practitioners validate compressed language models. While standard quality metrics like perplexity and downstream accuracy pass with flying colors, gently-compressed models systematically hallucinate procedural steps when deployed as agents, inventing instructions never present in the original task specification. The vulnerability appears specific to low-rank compression techniques like SVD truncation, not magnitude pruning. This disconnect between static evaluation and agentic behavior suggests current guardrails miss emergent failure modes that only surface under deployment, forcing teams to rethink compression validation pipelines before shipping production agents.

Modelwire context

Explainer

The paper's core finding isn't just that compression breaks agentic tasks, but that it breaks them in ways invisible to every standard quality gate. The hallucination is procedural and systematic, not random noise, suggesting compression damages reasoning scaffolding rather than general capability.

This connects directly to two concurrent threads in recent coverage. First, the Fairness Pruning work from late July showed that surgical model edits can target specific neuron populations without full retraining, implying that compression techniques similarly affect localized model behaviors in ways global metrics miss. Second, the CDAE robustness paper from the same period demonstrated that perturbations (including structural ones like pruning) can shift model behavior unpredictably even when aggregate performance holds steady. The current finding extends this: compression passes MMLU and perplexity but fails under the sequential decision-making demands of agentic execution, suggesting evaluation frameworks need task-specific stress tests, not just aggregate benchmarks.

If teams deploying SVD-compressed models in production agents report similar hallucination patterns within the next two quarters, this moves from a research curiosity to a validation crisis. Specifically, watch whether major model providers (Anthropic, OpenAI, Meta) add agentic stress tests to their compression validation pipelines before Q4 2026, or whether the field continues treating compression as a solved efficiency problem.

Coverage we drew on

Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsFidelity · LLMs · SVD · MMLU

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “Fidelity Is Not Safety: Gently-Compressed LLMs Pass Every Data-Free Quality Guard Yet Invent Procedure Steps in Agentic Execution”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.