← All posts

Why Sampling Temperature Does Not Eliminate LLM Fingerprints

Randomizing token selection doesn't remove the underlying signals that reveal AI model provenance.

October 31, 2026

Cover illustration for “Why Sampling Temperature Does Not Eliminate LLM Fingerprints”

Sampling temperature does not eliminate LLM fingerprints because it changes token selection without removing the statistical sources that make outputs recognizable. Training, architecture, watermarking, and the inference system leave signals that survive randomized sampling. To suppress fingerprints, a system must therefore change those sources or the signals they build up, while provenance compliance calls for documented records that a generation parameter cannot provide.

What temperature actually does to a model's logit distribution

Temperature is a sharpness control applied to raw logits before sampling. It scales their values uniformly, making the resulting probability distribution flatter at high temperature or more peaked at low temperature. For a given set of logits, positive temperature scaling preserves the ranked ordering of tokens. What changes is the strength of that preference and, consequently, how often the sampler selects a less probable option.

That distinction separates the model’s computed preferences from the sampling rule applied to them. Training data, architecture, and the inference system shape the raw logit landscape, and a change in temperature leaves all those sources in place. The resulting probability distribution does change, so describing the probabilities themselves as invariant would be inaccurate. The stable foundation is the process producing the logits. Temperature changes how the sampler uses that foundation.

The two distinct fingerprint signals that live in the logit landscape

Deliberate watermarks and intrinsic stylometric signatures occupy different parts of this process. In the KGW-family mechanism described here, a hash of the previous token partitions the vocabulary into a green list and a red list. A bias δ is then added to green-token logits after temperature scaling, favoring their selection. Randomness does not remove that directional preference. Higher entropy broadens the available choices without suppressing the expected overrepresentation of green tokens.

Detection uses a z-score to measure that overrepresentation by comparing the green-token count in the text with a threshold. As text length increases, the detector accumulates more evidence of the bias. Intrinsic signatures, by contrast, require no deliberately injected signal. These model-specific writing styles, termed aidiolects, emerge from training distributions, attention structure, and architectural priors. Their lexical, syntactic, and rhetorical preferences persist across topics and prompts, allowing detection from surface features without a secret key.

The two signals therefore resist temperature for different reasons. The watermark bias enters downstream of temperature in the mechanism described, while the sources of stylometric preferences precede the inference call. Both sources stay in place when tokens are chosen more randomly. This does not mean every detector has identical accuracy at every setting. Temperature can change how a signal appears in generated text while leaving its cause in place.

A third fingerprint layer temperature cannot see: the inference system itself

Wimbauer et al. (2026) find another signal that comes from the inference engine, attention backend, and hardware platform. Floating-point non-associativity produces tiny but systematic differences in computed logits when the same model runs on an Nvidia A100 versus an H100. Those differences propagate into observable text. The study can tell which system components were used from prompt-response pairs alone, without access to logits. Its method therefore does not depend on recovering a deliberate watermark or reading a model’s writing style.

The study reports perfect identification across tested components at temperature zero and 76–80% accuracy at non-zero temperatures across many configurations. The authors conclude that prevention is fundamentally hard because it would require eliminating numerical differences between hardware and software stacks. Those differences are outside the temperature parameter’s control. Even suppression of watermark and stylometric evidence would leave this third channel to address, further weakening temperature’s claim as a general defense.

What benchmark results show about fingerprint survival under temperature manipulation

Gloaguen et al. (ETH Zurich) report that semantically conditioned watermark fingerprints survive sampling variations within 25 deployment scenarios tested. Their 2025 study records a 100% detection rate across scenarios including finetuning, quantization, pruning, sampling variations, and active adversaries. The fingerprint is diffused statistically throughout responses in a predetermined semantic domain rather than confined to brittle, fixed responses. This is a different research program from inference-system identification. Nevertheless, its sampling results point in the same direction.

The studies do not measure an interchangeable detection task. One identifies inference components, while the other tests a model-ownership fingerprint, so their figures should not be treated as a single universal detection rate. Their shared result is narrower and useful. Even when sampling variation changes individual responses, enough structure remains for identification to work under the tested conditions.

Stylometric detection findings in the research brief extend that pattern to surface features, with reported accuracy in the high eighties to high nineties without a secret key. Across these studies, temperature is one of many changes tested, with no special power to defeat detection. The brief characterizes it as routinely the weakest attack vector. These results support the account of how temperature works without claiming that every fingerprint always survives.

Where temperature does have a partial effect, and why that partial effect degrades quality before it defeats detection

TempParaphraser (EMNLP 2025) finds that higher-temperature sampling improves distribution smoothness and makes detection harder, while also degrading output quality. Some detectors become less confident when outputs are less concentrated around familiar choices. Reduced confidence is not equivalent to crossing a detection threshold. The same intervention that weakens some recognizable patterns also worsens the text.

For KGW watermarks, the limitation is more specific. Raising temperature increases variance around the green-token rate without directly suppressing that rate. Longer sequences can therefore accumulate enough green-token evidence to exceed the detection threshold even under high-temperature sampling. In the tradeoff described by the brief, temperature high enough to meaningfully suppress a z-score produces noticeably worse content, making direct intervention in the green-token rate the relevant alternative.

Why effective watermark evasion requires targeting the green-token rate, not the temperature parameter

Bias Inversion Rewriting Attack, or BIRA, targets the statistic the watermark detector actually reads. Its rewriter suppresses the conditional probability of selecting a green token on average across a sequence. With this approach, even a small per-step margin can make detection probability fall exponentially. Instead of flattening the full distribution, the rewriter acts against the systematic excess of green-list selections.

BIRA demonstrates a reduction from a z-score well above the detection threshold to one well below it while preserving meaning. This result addresses the mechanism that temperature leaves intact. It does not establish that every rewriting system will achieve the same outcome, or that defeating a watermark removes the other fingerprint layers. The intervention is specific to the detector’s statistical signal. Its significance is the match between what the detector counts and what the rewriting process changes.

For content teams, the architectural implication is a change in the question being asked. Content teams need to assess whether the generation system disperses the statistical signal detectors seek, rather than which temperature setting makes prose least predictable. That shifts attention from a per-generation parameter to system design. It also calls for testing each signal on its own, because a watermark result alone cannot show that stylometric or inference-system evidence is gone.

Why multi-model architectures disperse the fingerprint where single-model temperature adjustments cannot

A single-model pipeline draws its output from one training distribution, one architecture, and one inference system, regardless of temperature. Changing the sources of generation can therefore change something that sampling adjustments leave in place. But using several models is not sufficient by itself. A sequence in which one model drafts, another refines, and another formats can leave overlapping or partially degraded watermarks and visible stylometric seams. Sequential chaining changes contributors without necessarily preventing their signals from accumulating.

The research brief notes that existing watermarking methods are designed for single-model settings and do not trace information flow across multiple agents. Token-level mixing across providers and architectures instead aims to prevent any single model’s green-list bias or stylometric signature from accumulating throughout the sequence. This is the structural premise of genuine dispersal.

Lettershred puts that principle to work with a dozen model kernels that draft in parallel across providers, a judge panel that selects output line by line, and a final enhance pass that gives the text one voice. The company reports a substantial watermark z-score reduction to below the detection threshold of 4, with roughly half of tokens rewritten. That is the company’s own quantitative claim. It illustrates the proposed architectural mechanism, but it is not independent evidence that every fingerprint channel has disappeared. A low watermark score alone also cannot show that a system meets regulatory requirements.

The compliance pressure that makes architectural fingerprint dispersal urgent rather than optional

The GPAI Code of Practice, published July 10, 2025, makes watermarking and provenance tracking monitoring obligations for signatories. The EU AI Act requires documented data provenance for high-risk AI systems under Article 10, with enforcement beginning August 2026. These obligations bring generation architecture and documentation into the same operational discussion. They do not make evasion equivalent to compliance. A system that disperses a detectable signal still needs the records required for its applicable obligations.

A positive watermark result indicates that AI generation was involved, but it does not attribute an artifact to an organization, model version, or timestamp. Provenance requires a structured record capturing the governing specification, model and provider, prompt or task, human who accepted the output, and tests that passed. Detection answers a narrower question than documentation. Treating temperature as deniability therefore leaves an organization without the information needed to explain how an artifact was produced.

C2PA provenance records are not retroactive. Content produced before provenance infrastructure was deployed has no such record, regardless of its generation temperature. Teams that rely on sampling changes thus risk a positive detection result with no documentation, instead of having a documented AI-generated artifact with a clean z-score. The practical implication is to address provenance as part of system design. The choice is whether to establish that record before or after an enforcement event.

The key objection to fingerprint persistence, and why it does not rehabilitate temperature as a defense

KGW detection depends on the secret hash key used to partition the vocabulary. Without that key, an external party cannot reconstruct which tokens were green-listed or compute the z-score for that watermark instance. Access is a real constraint on this detector. The output may still contain detectable evidence from other sources.

Intrinsic stylometric detection reads lexical and syntactic features without a watermark key. The inference-system study likewise identifies components from prompt-response pairs without logit access or knowledge of watermarking parameters. Changing temperature neither supplies a missing key nor removes the bias a properly keyed detector would measure.

Zhang et al. (2026) expose a different limitation through GhostPrint, which fine-tunes a weaker model to mimic a stronger model’s fingerprint. Their result shows that fingerprinting pipelines can be vulnerable to spoofing by a motivated actor, rather than proving fingerprints are immutable identifiers. That qualification is essential to interpreting detection evidence. The demonstrated intervention changes model behavior through fine-tuning, so temperature-based randomized sampling remains unsupported as a defense.

Fingerprint persistence is therefore not a claim of infallible attribution. Secret-key access limits one detector, and spoofing exposes weaknesses in others.

LLM Fingerprinting

Sources