TL;DR
Research reveals that LLM "fingerprints"—the stylistic traits used to identify which model wrote a text—are contagious. When a model continues a conversation started by another, it subconsciously mimics the predecessor's style, allowing dishonest API resellers to stealthily swap premium models for cheaper ones mid-conversation. To combat this, the authors introduce LINEUP, a robust auditing method that catches these "identity thefts" by comparing replies against references generated using the exact same conversation history.
What Problem Does This Paper Address?
We often assume that an LLM has a fixed "voice" or fingerprint that can be used to identify it. This assumption underpins many auditing tools designed to ensure that when you pay for a "frontier" model (like Claude 3.5 Sonnet), you aren't being served a cheaper substitute (like DeepSeek or a smaller "Haiku" version) by a shady API relay.
However, this paper identifies a critical flaw in current auditing: fingerprints are not static. In multi-turn conversations, a model's style is heavily influenced by the history of the chat. This creates two major problems for existing audits that rely on "fresh-context" (history-free) references:
- False Negatives (Camouflage): A cheap substitute continuing a premium model's conversation will start to sound like the premium model, "hiding" in its stylistic wake.
- False Positives (Drift): An honest premium model's style naturally drifts over a long conversation. If an auditor compares this drifted style to a "fresh" reference, it might wrongly flag the honest model as a substitute.
The economic incentive for this "mid-conversation substitution" is massive. In an 8-turn dialogue, the later turns (5–8) account for nearly 80% of the input tokens. Swapping to a cheaper model after turn 4 can save a provider over 50% on their bill while being nearly invisible to standard detectors.
Methods
The authors first quantify the "contagion" effect using three types of text classifiers: n-gram (word/character patterns), neural (LLM2Vec embeddings), and stylometric (43 specific features like punctuation and sentence length).
To fix the auditing problem, they propose LINEUP. Instead of using a pre-collected, static reference for a model, LINEUP works as follows:
- At an audited turn, the auditor takes the actual conversation history.
- The auditor sends this history to the "official" claimed model $m$ times to get $m$ reference replies.
- The auditor then ranks the relay's reply among these $m$ references using a scoring function (like the logit from a classifier).
- If the relay is honest, its reply should be indistinguishable from the references. If it's a substitute, it will likely stand out as an outlier in the "lineup."
Datasets
The study utilized a diverse set of 13 models, including frontier closed models (anonymized/future-dated in the paper as Claude Sonnet 5.5 and GPT-6.1 Sol) and various open-weight models like Gemma-3-27B, OLMo-3.1-32B, and Qwen3-32B.
The conversation data was built from:
- WildChat: 200 seed requests from real-world logs.
- Llama-3.1-8B Simulator: Used to generate consistent user turns so that the "user" behavior didn't change based on which assistant was replying.
- Confirmation Set: 60 additional seeds focused on specialized tasks like coding, math, and factual QA.
Key Insights and Ideas
- The Contagion is Stylistic, Not Verbatim: The models aren't just copying words from the history; they are adopting the style. The authors proved this by restating the history in a different style (using a third-party paraphraser), which significantly reduced the contagion effect.
- History is a Strong Confounder: A model's fingerprint is a property of the model and its history. This means any audit that doesn't account for history is fundamentally flawed.
- Active Probes are Brittle: Even simple "identity" probes (e.g., asking a model to "flip a coin" to test its randomness) fail on long contexts. Honest models were flagged as "dishonest" up to 100% of the time simply because their internal distributions shifted over the course of the chat.
Key Results
- Detection Collapse: For a passive classifier, the ability to detect a substitute drops from a near-perfect 0.97 to 0.72 when DeepSeek replaces Claude Sonnet mid-chat. For same-vendor downgrades (Sonnet to Haiku), it drops from 0.87 to 0.54.
- LINEUP Success: LINEUP maintains a low false-positive rate of 0.056 (meeting its 0.10 preregistered bound) and achieves a detection rate of 0.81 after just four audited turns across six different model pairs.
- MMD Variant: A more powerful version of the test using Maximum Mean Discrepancy (LINEUP-MMD) was able to detect 100% of substitute sessions.
Limitations and Future Work
- Cost of Auditing: LINEUP requires multiple calls to the official (expensive) API to generate references for every audited turn.
- Score-Aware Attackers: If a relay knows the specific scoring function an auditor uses, they could perform "best-of-n" sampling to pick the substitute reply that looks most like the claimed model. The authors suggest keeping the scoring function private or using multi-dimensional statistics (like embeddings) to defend against this.
- Stationarity: The method assumes the official model doesn't change significantly between the time the relay replies and the time the auditor generates references. "Contemporaneous" sampling is required.
- Scope: The study focused primarily on chat; future work is needed to see how these effects manifest in more complex "agentic" workflows involving tool use and long system prompts.
Summary updated