Does Slightly Mean Somewhat? Measuring Vague Intensity Words in LLM Numeric Actions

2026-05-22 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

A study conducted in May 2026 investigated how Claude Haiku (claude-haiku-4-5-20251001) translates vague intensity words into numeric actions within a controlled resource-allocation environment. Across 6,620 runs at T=0.0 and T=0.7, the model compressed 10 intensity words into 5 distinct median outputs, with four lower-tier words mapping to a 0.50 allocation (Spearman ρ=0.845, p<0.001). When the current system state was provided, starting allocation dominated word choice, capturing ε²₋ₐₓₗₑₔ=0.782 of rank-based variance compared to ε²ₗₒₓ₊=0.079 for word grouping. Near feasibility limits, the model exhibited three behavioral modes: weak words made small adjustments, strong words often abstained, and "drastically" pushed to the local ceiling of 0.912. These patterns remained consistent across temperature settings, broadening distributions but not restoring ordinal distinctions.

Key takeaway

For AI Engineers designing natural-language control systems, understand that LLMs like Claude Haiku do not interpret vague intensity words as stable, scalar actions. Your interfaces must account for significant compression of lexical distinctions and strong state-dependence, where context overrides word meaning. Test system behavior near operational limits, as small lexical changes can cause unexpected abstention or ceiling-pushing actions, leading to unreliable system responses.

Key insights

LLMs compress vague intensity words into fewer numeric actions, heavily influenced by system state and exhibiting discontinuous boundary behavior.

Principles

• LLM numeric actions compress fine-grained lexical distinctions.
• Context state dominates word choice in LLM numeric actions.
• LLMs exhibit discontinuous action modes near operational limits.

Method

A controlled experiment using a synthetic resource-allocation environment, feeding 10 intensity words to Claude Haiku to produce numeric tool calls, then evaluating outcomes.

In practice

• Avoid relying on fine-grained distinctions in vague LLM instructions.
• Explicitly provide system state when using vague modifiers.
• Test LLM behavior near operational limits for critical actions.

Topics

• Large Language Models
• Natural Language Interfaces
• Vague Language
• Numeric Actions
• Claude Haiku
• Prompt Sensitivity

Best for: Machine Learning Engineer, NLP Engineer, Research Scientist, AI Scientist, Prompt Engineer, AI Engineer

Related on AIssential

• Unveiling the Limits of Large Language Models in Inferring Pragmatic Meaning from Non-Verbal Responses — LLMs significantly struggle to infer pragmatic meaning from non-verbal communication, unlike their performance with verbal cues.
• How Aara Reads: The Secret Language Beneath the Words — LLMs translate human language into numerical vectors via tokenization, embedding, and attention to process meaning.
• the magic words to master AI — Specific, well-crafted prompts significantly improve AI output quality and utility.
• Finite Certificates for In-Context Determinacy and a Threshold Theory of Emergence in Language Models — The paper provides a model-theoretic framework using finite semantic certificates to verify language model behavior and understand emergence.
• Modeling Pathology-Like Behavioral Patterns in Language Models Through Behavioral Fine-Tuning — Consistent behavioral fine-tuning can induce stable, pathology-like biases in LLM generative distributions and action selection.
• Why Temperature 0 Doesn’t Make Language Models Think — LLMs generate text via sequential token prediction based on conditional probability, not internal planning or reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.