acceptodds
Under review as a conference paper at ICLR 2027

Constitution-Guided Watermarking

Abstract

Watermarking enables language model providers to identify text generated by their models. However, its desired properties can conflict (i.e.,  stronger watermark signals can degrade text quality), while designs that resist editing may also facilitate forgery. Providers address these trade-offs by choosing configurations that balance competing objectives or prioritize particular properties. Either approach imposes a shared operating point on requests with different requirements, potentially sacrificing quality where wording preservation matters or robustness where reliable attribution is essential. To allow flexible and adaptable designs, we introduce Constitution-Guided Watermarking, a framework that selects request-appropriate trade-offs from provider requirements, listed as natural-language principles. Offline, a pretrained reasoning agent examines constitutional rules alongside watermark implementations and iteratively refines rule-specific configurations using empirical feedback. At deployment, a separate monitor identifies applicable rules and retrieves the corresponding policy, including watermarking exemptions, without modifying the serving model. Furthermore, our framework supports offline parallel optimization and refinement of rule-specific configurations based on evolving provider requirements without affecting deployment, and binds each deployed configuration to its evaluation evidence, making deployment decisions auditable. In a proof-of-concept evaluation using KGW and a five-rule constitution, our framework selects configurations responsive to provider priorities and improves post-paraphrase detection on robustness-prioritized requests by up to percentage points over fixed configurations, while matching or exceeding all baselines in aggregate quality and clean detection at a nominal false-positive rate.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large Language Models, Watermarking, Constitutional AI

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HMXcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a Constitution-Guided Watermarking framework: service providers write a watermark constitution in natural language, a verification agent searches for and iteratively refines watermark configurations for each rule, and at inference time a deployment monitor identifies which rules apply to a request and retrieves and executes the corresponding policy, while the service model remains frozen; each policy is recorded as a certificate documenting the configuration, measured evidence, and unmet criteria.

Critical strengths

The paper proposes a mechanism in which a verification agent reasons over natural-language constitution rules and adjusts watermark parameters. This is a distinctive contribution to configuration selection for text watermarking: it turns what properties the provider wants to prioritize from hyperparameter tuning into an auditable rule-based input, with the agent iteratively proposing configurations on an offline reference set.

Critical weaknesses

The paper aims to show that constitution guidance can improve robustness on a per-rule basis, but the absolute level of robustness is far below the thresholds the authors themselves set, and the relative improvement rests on a small-sample search of only 10 reference prompts per rule; moreover, the FPR is not an independent audit and cannot support the strong conclusion that "the policy is effective." Second, the evaluation scope and baselines are insufficient to isolate the contribution: a single watermark family, a single model, a single key, and a single seed, with comparisons only against fixed configurations and random search, and no Bayesian optimization, hand-crafted policies, or specialist-trained agents.

Questions for authors

This paper claim to address "constitution-guided watermark policy selection," yet the experiments use only a single base watermark method (KGW), and the verification agent searches only over KGW's parameters. This makes it difficult for me to judge whether the framework's conclusion is that "the constitution mechanism is effective" or merely that "KGW happens to find a relatively good point within this small configuration space." I suggest you instantiate the framework on at least multiple watermark families and have the agent make choices at the algorithm level, rather than only tuning KGW hyperparameters.

Reviewer_oLxNOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper deals with a real deployment problem for text watermarking: different requests need different watermarking behavior, and no single global setting can satisfy all of them. The authors propose Constitution-Guided Watermarking, where the provider writes a constitution with ordered rules, each with a priority and a quantitative acceptance gate. An offline agent proposes and tunes a watermark configuration for each rule using reference-set feedback, and an online monitor routes each incoming prompt to the right policy. The underlying model stays frozen. Every deployed configuration comes with its evidence and is labeled either REFERENCE-PASSED or BEST-EFFORT depending on whether it cleared its gates.

As a proof of concept, they instantiate this with one LLM as the verification agent, another as the deployment monitor, and KGW as the watermark. They test five rules, covering things like misuse-prone requests, low-risk editing, high-stakes advice, deterministic answers, and general generation. Compared to two fixed KGW presets and a random search baseline, Constitutional-KGW gets about 98% TPR at 0.1% FPR, a 98.6% perplexity-based pass rate, and correctly skips watermarking for editing requests. On the rule that needs the most robustness, it pushes post-paraphrase detection up to 13.8% and 17.0% against two paraphrasers, versus 2.8% and 3.0% for one of the fixed presets. That said, the paper's own 50% robustness target still isn't met on that rule, so it ends up labeled BEST-EFFORT instead of REFERENCE-PASSED.

Critical strengths

Strength 1 — I like that every deployed configuration comes with its evidence attached. The paper sets up the constitution and the CONFIG operator so that each deployed policy ships together with its evidence and the list of gates it failed, including the BEST-EFFORT labels. That means the tables aren't just results, they're a policy plus an audit trail. For anyone who has to actually audit a deployed system, that's genuinely useful.

Strength 2 — The paper is honest about where it comes up short. It states plainly that paraphrase robustness stays below the 50% provider gate, and it pairs the misuse rule with the gates it didn't clear. It also reports routing failures and at least one configuration that got archived instead of deployed. Showing the messy, partial state instead of only the wins is exactly the kind of disclosure I want to see more of.

Critical weaknesses

Weakness 1 — The headline "~14-point" robustness gain looks shakier once you think about what it's being compared to. Fixed KGW A is just the off-the-shelf setting, never tuned for this task, so of course a tuned configuration beats it. The random-search baseline, which is actually tuned, gets much smaller gains, and the confidence interval includes zero for most comparisons. On top of that, the misuse-rule configuration the paper ends up with is just a stronger-bias KGW setting overall. The obvious ablation is to test that exact same stronger setting without the constitution or the agent, and see if you get similar robustness. That ablation isn't in the paper, so I can't tell how much of the gain comes from the framework versus just picking a stronger fixed point.

Weakness 2 — The rule that needs robustness most doesn't actually meet the paper's own 50% gate. The intro argues we need something adjustable because fixed points fail to satisfy every request, but then the one rule where that matters most still falls short even with the framework. Labeling it BEST-EFFORT is honest, but it does mean the core contribution is narrower than the intro suggests.

Weakness 3 — The false-positive numbers aren't as independent as they look. The same null-text sample used during configuration search also gets reused to report the false-positive rate. The paper admits this in an appendix, which I appreciate, but the main text and abstract still read as if the FPR numbers were independently measured. I'd want that caveat moved up front where the main numbers are reported.

Questions for authors

A. Questions that could change my assessment.

  1. The misuse-rule configuration ends up at a much stronger-bias KGW setting than the fixed baseline. Could you add an ablation that tests this exact stronger setting without the constitution or the verification agent? That would tell us how much of the robustness gain is from the framework versus just a better fixed choice.

  2. The paper mentions the tested constitution only uses first-applicable rule precedence, so the more general rule-combination logic in the framework never actually gets exercised. How much of the claimed generality depends on that unused logic, versus policies that just pick one matching rule and stop?

  3. There's a theoretical extension for forgery resistance that assumes certain conditions the paper doesn't test. How would a provider actually check those assumptions hold in a real deployment, and how much of that security claim is still just on paper?

  4. The false-positive numbers get calibrated and reported on the same null-text sample. Could you report an independent false-positive audit on a fresh, held-out null set? If that's not feasible right now, could the abstract and main results at least flag that the reported numbers aren't from an independent audit?

  5. The test prompts and their paraphrases are synthetically generated. Could you report some independent human validation of a sample, just to confirm the rule labels and paraphrase quality hold up outside the synthetic pipeline?

B. Suggestions for future work.

  1. Add the fixed-configuration-without-agent ablation mentioned above, holding the search budget and watermark family constant.

  2. Run an independent false-positive audit on data that was never touched during configuration selection.

  3. Compare the constitutional routing against something simpler, like a prompt classifier, to see how much the extra machinery actually buys you.

  4. Try this with more watermark families, more serving models, and more constitutions to see how well it generalizes.

Reviewer_4HtgOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework to guide watermarking using rules in order to adjust the trade-offs between text utility versus detectability and robustness of the watermark, depending on the type of prompt (e.g. a question about a medical document should prioritize text utility). The proposed method relies on an (offline) agent to determine the appropriate rules to apply, and an (online) monitor to select the rules for the current prompt. An evaluation of the framework is presented based on the KGW watermarking technique.

The goal is to tailor the quality of the model output for a provider wishing to apply watermarking while simultaneously prioritizing the usefulness of the responses for the user.

Critical strengths
  • The problem addressed is relevant and timely, as it answers a practical limitation of uniform watermarking policies. The paper is convincing that Indeed different prompts may require different tradeoffs between text utility and watermark quality

  • The proposed framework does not modify the serving model

  • The sensitivity experiments show that changing constitutional priorities leads to different configurations and shifts in measured outcomes in the intended direction, although the distinction between neighboring priorities remains limited.

Critical weaknesses
  • The results of the experiments show limited Added value of the proposal. In Table 8-9, the robustness improvement versus a random configuration search excludes zero in one of the four attack–FPR combinations (DIPPER at 0.1% FPR). While this provides evidence of an advantage, the consistency remains unclear. It seems that a simple hand made per-rule policy would further help isolate the contribution of the agent

  • The evaluation only considers KGW configurations, with one serving model and one search seed. Comparisons with other existing watermarking methods would strengthen the generality of the proposal

  • The global perplexity threshold might be confusing between the domain specific language and the degradation due to the watermark. For example, specialized medical answers could exceed the threshold even without watermarking. Reporting unwatermarked pass rates and paired changes per scenario would help disentangle these effects. It might be useful to complement PPL with distributional divergence or semantic embedding distance to assess distortion, alongside task-based utility measures.

  • The formalization is heavy compared with the evaluation, and it would benefit from explicit links to the experimental metrics. The main text should also more clearly distinguish the contribution from prior adaptive watermarking and separate evaluated components from proposed extensions

Questions for authors
  • How does the reasoning agent compare with a simple "hand-made" policy assigning stronger watermark to misuse-prone requests and weaker configurations elsewhere? Do the gains over random search persist across multiple search seeds?

  • Can the framework adapt and demonstrate similar benefits with a watermarking algorithm other than KGW?

  • What are the unwatermarked PPL pass rates for each scenario? Could the global threshold penalize specialized answers independently of watermarking? Why not complement PPL with paired distortion measurements and task-specific utility evaluations?

  • Do the selected configurations maintain their nominal false-positive rates on fresh null texts that were never used during calibration or configuration selection?

  • Does the agent retain an advantage over random search when matching total computational cost rather than only the number of evaluated configurations?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.