The slain cliché: how one random word can kill a story — and resurrect your model’s hidden map

2026-03-10 · Source: Data Science on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, long

Summary

The article introduces "concept seeding," a method developed at Stanford to analyze how Large Language Models (LLMs) process information by injecting a single, randomly chosen word into a prompt. This technique reveals the model's internal map of meaning and its stability under pressure. Experiments showed that a single seed word, like "rickshaw," could override extensive context (e.g., 197,000 characters of Shakespeare) even at temperature=0. The method involves sampling a random point in an embedding space, finding the nearest real word, and prepending it as an "inspiration word" to the prompt. This approach generates significantly more diverse "random" words (e.g., 8,347 unique words from 10,000 draws) compared to asking the model to generate a random word itself (1,247 unique words). Concept seeding also helps identify "hidden invariants" by showing which claims remain stable (high certainty scores) and which are fragile (low certainty scores) across multiple seeded runs, as demonstrated with factual questions on Gemini models.

Key takeaway

For research scientists evaluating LLM behavior or seeking to understand model knowledge structures, you should integrate concept seeding into your analysis workflow. This method allows you to probe the epistemic stability of claims, distinguishing deeply internalized information from speculative or peripheral statements. By observing what persists and what changes under diverse conceptual perturbations, you can gain a clearer picture of a model's internal landscape, informing more robust model evaluation and application strategies.

Key insights

Concept seeding reveals LLM internal knowledge maps and claim stability by perturbing prompts with single, geometrically sampled words.

Principles

• Context is not always dominant over a single, strong concept seed.
• LLMs exhibit internal rigidity on factual prompts, but creativity on open-ended ones.
• Epistemic stability (consistency) is distinct from factual correctness.

Method

Sample a random point in an embedding space, find the nearest real word, and prepend it as an "inspiration word" to an LLM prompt. Repeat with different seeds to observe output stability.

In practice

• Use concept seeding to break LLM repetition in creative tasks.
• Identify stable vs. fragile claims in LLM outputs for research.
• Detect shifts in model knowledge between different LLM versions.

Topics

• Concept Seeding
• LLM Evaluation
• Epistemic Stability
• Embedding Spaces
• Model Behavior Analysis

Code references

• ClancyDennis/concept-seed

Best for: Research Scientist, AI Researcher, AI Scientist, Machine Learning Engineer

Related on AIssential

• Research on LLM alignment as latent discourse-level regimes vs. token-level filtering? — LLM alignment may stem from global latent space geometry and discourse-level regimes, not just token-level filtering.
• Accuracy, Stability, and Repeated-Run Reliability of Large Language Models on Deterministic Programming Tasks — LLM evaluation for deterministic tasks requires repeated-run stability analysis to complement conventional single-run accuracy reporting.
• Does Self-Consistency Improve the Recall of Encyclopedic Knowledge? — Self-consistency improves both symbolic reasoning and encyclopedic knowledge recall, even when CoT primarily aids reasoning.
• Faithfulness Serum: Mitigating the Faithfulness Gap in Textual Explanations of LLM Decisions via Attribution Guidance — LLM-generated explanations often lack epistemic faithfulness, but a new method improves it via attention-level guidance.
• Navigating Unreliable Parametric and Contextual Knowledge: Explicit Knowledge Conflict Resolution for LLM Inference — MACR resolves LLM knowledge conflicts by explicitly assessing confidence in sources and employing multi-agent reasoning, moving beyond binary reliability assumptions.
• Context Graphs: Free Association, Consistency, Narratives & Truth — LLMs function via free association on narrative graphs, trading consistency for broader, stochastically generated outputs.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Data Science on Medium.