Auditing Cross-Lingual Fairness in Language Model Watermarking
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Auditing Cross-Lingual Fairness in Language Model Watermarking".
Jane: Here's the gist of the paper. Watermarking schemes for large language model output are evaluated almost exclusively on English text using each scheme’s detection threshold and a narrow set of quality measurements on English text using each scheme’s detection threshold and a narrow set of quality measurements > Cross-language disparity is large, scheme-specific, and predominantly between-family on the typological partition rather than idiosyncratic to particular languages on both detection and quality > Evaluation Framework Components The authors propose an evaluation framework with four components designed for cross-lingual settings empirically calibrated detection thresholds, a threshold-independent companion measurement that distinguishes calibration failures from detection failures, three disjoint quality measurement paradigms (distributional, paired-semantic, and reference-perplexity), and a generalized-entropy decomposition of cross-language disparity over a pre-registered typological partition.
Tom: Audit Scope and Findings The audit was applied to six watermarking schemes, three open-weight generators, eleven languages spanning four scripts and eight typological families and both base and instruction-tuned regimes, yielding ≈ 200,000 matched watermarked/unwatermarked generations on shared prompts.
Jane: The framework reveals failure modes that single-language single-paradigm evaluation cannot surface.
Tom: Across detection and quality, observed disparity is predominantly between-family on the typological partition rather than idiosyncratic to particular languages.
Title and authors: Tom: So we're looking at this paper titled "Auditing Cross-Lingual Fairness in Language Model Watermarking." It sounds like they are finally taking a serious look at how these watermarking schemes work across different languages.
Jane: Yeah, that’s right. The authors are essentially saying that most of the research on watermarking has been done only on English text, and that assumption is actually a big problem when you think about real-world use with many languages.
Lu: Exactly. They propose this whole evaluation framework to treat cross-lingual fairness as a main thing to measure instead of just doing an after-the-fact check on English benchmarks.
Meng: I'm curious how they’re going to test this across so many different things—six schemes, three generators, and eleven languages spanning four scripts and eight typological families? That sounds like a massive experimental setup.
Lalam: The grid they built is quite detailed; they are testing every cell of an eleven languages times six schemes times three generators times two regimes grid at n equals five hundred matched pairs per cell on shared prompts.
Tom: Okay, so what’s the core idea here? What did they actually measure to find the fairness gaps?
Jane: The main point is that the disparity they found in detection and quality isn't random; it's mostly structural, meaning it relates to language properties like script or grammar rather than being specific to one language.
Lu: They used a generalized-entropy decomposition of disparity over a pre-registered typological partition to see if the gaps were structural or just random noise in the data.
Meng: If it’s structural, that means some languages are inherently easier or harder to protect, regardless of the specific watermark they use. That’s something we need to know when we talk about real deployment costs and fairness at scale.
Lalam: The results showed that for cells where the generalized entropy number was above the floor of one to three, the share between different language families was high, sitting between sixty-one point three percent and ninety-nine point nine percent.
Tom: That sounds like a lot of variation in fairness across those families. And what about detection failures versus calibration failures?
Jane: They introduced a threshold-independent companion measurement to separate these two things. For instance, they showed that you can distinguish a calibration failure where the true positive rate is near zero but the area under the curve is high from a pure detection failure where both are low.
Title and authors: Lu: That’s smart because it helps pinpoint *why* something failed—whether it’s the detector not being tuned correctly or if the watermark itself isn't working well in that context.
Meng: And they also looked at how instruction tuning affects things, which is crucial for how models are actually used by people.
Lalam: They found that instruction tuning causes a collapse in detection performance, with the mean true positive rate below zero point seven in sixteen of eighteen cells and a drop of roughly zero point one seven to zero point seven eight in the mean true positive rate compared to their base versions for non-SynthID cells.
Tom: So, even with instruction tuning, the detection can actually get worse on average for some parts of the test set. That’s a bit concerning when you’re trying to build robust systems.
Jane: It shows that we have to be careful about how we deploy these models; instruction tuning seems to swap the ranking of performance in certain areas, which is something we need to watch out for.
Lu: On the quality side, they used three separate paradigms: distributional, paired-semantic, and reference-perplexity. This is important because no single metric can tell you if a scheme is actually preserving the text well across all possible distortion mechanisms.
Meng: I've seen that in my startup work; you need multiple metrics to make sure something isn't just fooling one type of test while failing another.
Lalam: The paper found that distortion-free schemes like EXPEdit and SynthID were empirically the most distorting under every single one of those three quality paradigms, which is a bit surprising.
Tom: That means even the best looking schemes have some inherent weakness when it comes to preserving text fidelity. What about how instruction tuning affects quality?
Jane: Interestingly, instruction tuning causes a compression upward in all three of those measurements. For example, SynthID on Gemma instruct showed near-perfect quality preservation, which suggests the watermark has less effect on the generation distribution in that specific case.
Lu: That’s an important detail because it shows that some schemes are more robust to that kind of tuning than others, which is a finding we need to keep in mind as we design new methods.
Title and authors: Meng: From a deployment standpoint, if one scheme performs better under instruction tuning, that might be the one I lean towards for production use, assuming the cost of implementation isn't too high.
Lalam: The structural disparity analysis confirmed what they suspected: when the generalized entropy number is above zero point three, the between-family share is predominantly driven by structural factors like script or morphology rather than language quirks.
Tom: So this paper really shifts our thinking from just testing on English to understanding how fairness gaps are built into the very structure of language itself. What’s the practical implication for us right now?
Jane: The implication is that we can't treat watermarking as a one-size-fits-all solution; different languages require different levels of protection because their underlying linguistic structures dictate where those vulnerabilities lie.
Lu: It means the same scheme might offer materially weaker protection to speakers of some language families than others, and the cost of being watermarked isn't evenly distributed across all users.
Meng: That tells me that if we’re building a commercial product, we have to segment our deployment based on these structural risks instead of just picking the highest scoring scheme generally.
Lalam: The authors are proposing this four-component evaluation framework—calibrated thresholds, the companion measurement, three quality paradigms, and the entropy decomposition—as the first systematic way to measure this across languages.
Tom: So to wrap up what we've heard about "Auditing Cross-Lingual Fairness in Language Model Watermarking," it’s a big methodological step that moves us away from English-only testing.
Jane: It gives us a concrete way to see exactly where the fairness problems are hiding, separating calibration issues from detection failures, and seeing if the disparity is real structural or just noise.
Lu: The main thing is that it forces researchers to consider multilingual fairness as a first-class object of measurement when they design these tools.
Meng: For my team, this means we need to stop focusing on one language benchmark and start looking at the typology of the languages we actually ship models for.
Lalam: The paper’s contribution is that it surfaces failure modes that you absolutely cannot find through single-language or single-paradigm evaluation alone.
Tom: It sets a new standard for how we should be testing any text manipulation technique in large language models moving forward. We'll be keeping an eye on this framework as we look at other papers.
The paper's summary: Tom: So to wrap up what we’ve heard about auditing cross-lingual fairness in language model watermarking, this paper introduces a new way to test these schemes across different languages on a systematic scale.
Jane: Right, it’s about moving past just looking at English text and finally getting a real look at how these watermarking methods behave when you use them with many different languages.
Tom: The core of it is this four-part evaluation framework they propose to measure everything from the detection threshold calibration to the actual quality of the text preservation across different language types.
Jane: And what’s really interesting is that they found the disparity isn't random; it’s mostly structural, meaning it relates to how a language is built—like its script or grammar—rather than being specific to just one particular language.
Tom: That means some languages are inherently easier or harder to protect with these watermarks, which changes how we think about fairness in AI deployment.
Jane: Exactly, and they did this by using a generalized entropy decomposition over a typological partition to quantify whether those fairness gaps are structural or just random noise in the data.
Tom: They also introduced this threshold-independent companion measurement to separate genuine calibration failures—where the detector is just wrong—from actual detection failures where the watermark itself isn't working as expected.
Jane: And they found that instruction tuning, which is how most people actually use these models, actually causes a collapse in performance for detection on average across many of those language groups.
Tom: That’s a bit of a headache for deployment because it means you can’t just rely on the base performance metrics when you start fine-tuning models for specific tasks.
Jane: It shows that we have to be much more careful about how we deploy these models, because the protection level isn't consistent across different language families.
Tom: The authors found that distortion-free schemes like EXPEdit and SynthID are actually the most distorting under every quality test, which is a bit surprising if you think they’re built to be robust.
Jane: It means we can’t just look at one metric for quality; you need to use several—distributional, paired-semantic, and reference-perplexity—to get a full picture of how well the text is actually being preserved.
Tom: So this paper isn't just about finding a better watermark; it’s about building a proper testing system that acknowledges the complexity of real-world multilingual environments.
Jane: It really pushes researchers to think about these fairness gaps as being built into the very structure of language itself, which is a big shift from just testing on English.
Tom: The practical implication here is that we can't treat watermarking as a one-size-fits-all solution; different languages require different levels of protection because their underlying linguistic structures dictate where those vulnerabilities lie.
Jane: It means the same scheme might offer materially weaker protection to speakers of some language families than others, and the cost of being watermarked isn't evenly distributed across all users.
Tom: We need to stop focusing on one language benchmark and start looking at the typology of the languages we actually ship models for when we talk about security.
Jane: The authors are proposing this four-component framework—calibrated thresholds, that companion measurement, three quality paradigms, and that entropy decomposition—as the first systematic way to measure this across languages.
Tom: So it’s a lot of methodology packed into one paper to show how seriously cross-lingual fairness needs to be taken in this field.
Jane: It sets a new standard for how we should be testing any text manipulation technique in large language models moving forward, by showing where the actual failure modes are hiding.
Tom: And next, we’re going to look at what they found about instruction tuning specifically and how that affects these detection scores.
The paper's improvements: Tom: So we’ve seen how they evaluated these schemes, and now they suggest some ways to make that framework even better for real deployment scenarios across different languages.
Jane: Right, it’s about taking those findings and proposing actual improvements so that the detection thresholds are tailored more precisely to where you actually plan on using the model.
Tom: They suggest empirically calibrating detection thresholds per deployment context by looking at quantiles from unwatermarked text, which is basically a way to model deployment with a separate detector for each language or one big global threshold.
Jane: That’s smart because it moves away from just picking one fixed number and lets you adjust the sensitivity based on the specific environment.
Tom: Then they introduce this threshold independent companion measurement, AUC c equals the probability of a true positive score being greater than zero, estimated using the Mann-Whitney U statistic.
Jane: That’s a concrete way to separate calibration failures—where your detector is fundamentally misaligned—from detection failures where the watermark just isn't working well in that moment.
Tom: And for quality, they emphasize using those three disjoint paradigms again, distributional, paired-semantic, and reference-perplexity because no single metric can tell you if the text is actually being preserved across different distortion mechanisms.
Jane: They also suggest designing systems where detection and quality are balanced by analyzing the regime spread in instruction tuning to see how it affects things.
Tom: They’re pointing out that for instruction tuning, we see a compression upward in all three measurements, which they argue is predominantly a calibration failure rather than a real change in quality preservation.
Jane: That tells us that sometimes what looks like the watermark losing quality is actually just the detector getting confused by the new tuning regime.
Tom: On the structural side, they want to develop models that can identify which specific variables—like tokenizer fertility or language adherence rate—actually drive detection performance and quality preservation on a per-language basis.
Jane: That would be incredibly useful for figuring out exactly which linguistic features are causing the most trouble for certain languages.
Tom: They also suggest designing instruction tuning regimes in a way that might compress the scores together but not erase their separation, aiming to keep the distinction between watermarked and unwatermarked text alive.
Jane: So these improvements focus on making sure our tools are actually calibrated for real use and that we understand exactly what linguistic features are causing the biggest problems across different languages.
Tom: It’s about moving from just reporting what happened to having a toolkit that helps us diagnose *why* it happened in the first place.
Jane: And that leads us right into how these structural findings translate into real-world fairness gaps we need to consider as we build systems at scale.
Conclusion: Tom: So to wrap up this paper, "Auditing Cross-Lingual Fairness in Language Model Watermarking," it’s really about giving us a proper, systematic way to test these schemes across different languages that goes beyond just checking English text.
Jane: It shows that the fairness gaps we see aren't random; they’re tied to the structure of the language itself, and they introduce this four-part framework for measurement.
Tom: The main implication is that we can’t treat watermarking as a one-size-fits-all solution; different languages require different levels of protection because their underlying linguistic structures dictate where those vulnerabilities lie.
Jane: It means the same scheme might offer materially weaker protection to speakers of some language families than others, and the cost of being watermarked isn't evenly distributed across all users.
Tom: The authors are proving that single-language testing just doesn't cut it when you’re talking about global fairness in AI systems.
Jane: And they did this by showing how detection failures can be separated from calibration failures, and how quality preservation works differently depending on the paradigm you use to measure it.
Tom: They really lay out the mechanics of what’s going wrong at a very concrete level, which is great for us engineers trying to build these systems robustly.
Jane: It gives us a way to see exactly where the failure modes are hiding, separating calibration issues from detection failures, and seeing if the disparity is structural or just noise.
Tom: This whole paper sets a new standard for how we should be testing any text manipulation technique in large language models moving forward.
Jane: It’s about moving past just looking at English benchmarks and forcing researchers to consider multilingual fairness as a first-class object of measurement when they design these tools.
Tom: We need to stop focusing on one language benchmark and start looking at the typology of the languages we actually ship models for when we talk about security.
Jane: The paper’s contribution is that it surfaces failure modes that you absolutely cannot find through single-language or single-paradigm evaluation alone.
Tom: So this work on "Auditing Cross-Lingual Fairness in Language Model Watermarking" is a big methodological step that moves us away from English-only testing.
Jane: It’s about giving us a concrete way to see exactly where the fairness problems are hiding, separating calibration issues from detection failures, and seeing if the disparity is structural or just noise.
Tom: And next, we’re going to look at what they found about instruction tuning specifically and how that affects these detection scores.
Jane: That’s right, because understanding that tuning effect is key to knowing how stable our watermarking protection actually is during a model's life cycle.
Tom: We'll keep digging into those results next, so stick with us.
Alexander Nemecek, Osama Zafar, Debargha Ganguly, Vikash Singh, Vipin Chaudhary, Erman Ayday
Case Western Reserve University
cs.CL, cs.CR, cs.LG
Submitted: 2026-08-20
Updated: 2026-11-10
Comments: Accepted at NeurIPS 2026 Evaluations & Datasets Track
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 83/100
The gist: Watermarking schemes for large language model output are evaluated almost exclusively on English text using each scheme’s detection threshold and a narrow set of quality measurements on English…
Key concepts
- Four-Component Evaluation Framework
- A comprehensive system designed to test watermarking schemes across multiple dimensions. It includes empirical FPR calibration, a threshold-independent companion measurement, three distinct quality metrics (distributional, paired-semantic, reference-perplexity), and a generalized entropy decomposition over language types.
- Calibration Failure vs. Detection Failure
- A distinction made using a threshold-independent companion measurement (AUCc). Calibration failure occurs when the model's performance metric is wrong but detection is poor (TPR≈0, AUC≫0.5). Detection failure occurs when the model performs poorly regardless of calibration (TPR≈0, AUC≈0.5).
- Quality Preservation Paradigms
- Three different ways to measure if the watermark damages text quality: distributional (MAUVE), paired-semantic (BERTScore), and reference-perplexity. These are used together because no single metric can accurately judge how distortion affects the generated text under various conditions.
- Generalized Entropy Decomposition
- A statistical tool that breaks down cross-lingual performance gaps based on language typological partitions (like script or family). This reveals whether fairness issues are universal structural properties of language types or specific, idiosyncratic flaws in certain languages.
Terminology
Summary
Watermarking schemes for large language model output are evaluated almost exclusively on English text using each scheme’s detection threshold and a narrow set of quality measurements on English text using each scheme’s detection threshold and a narrow set of quality measurements >
Cross-language disparity is large, scheme-specific, and predominantly between-family on the typological partition rather than idiosyncratic to particular languages on both detection and quality >
Evaluation Framework Components
The authors propose an evaluation framework with four components designed for cross-lingual settings empirically calibrated detection thresholds, a threshold-independent companion measurement that distinguishes calibration failures from detection failures, three disjoint quality measurement paradigms (distributional, paired-semantic, and reference-perplexity), and a generalized-entropy decomposition of cross-language disparity over a pre-registered typological partition.
Audit Scope and Findings
The audit was applied to six watermarking schemes, three open-weight generators, eleven languages spanning four scripts and eight typological families and both base and instruction-tuned regimes, yielding ≈ 200,000 matched watermarked/unwatermarked generations on shared prompts. The framework reveals failure modes that single-language single-paradigm evaluation cannot surface. p. 0 We
propose an evaluation framework with four components: detection thresholds
calibrated empirically per deployment context, a threshold-independent companion measurement that distinguishes calibration failures from detection failures,
three disjoint quality measurement paradigms (distributional, paired-semantic, and
reference-perplexity), and a generalized-entropy decomposition of cross-language
disparity over a typological family partition. Applied to six watermarking schemes,
three open-weight generators, eleven languages spanning four scripts and eight
typological families, and both base and instruction-tuned regimes, the framework reveals failure modes that single-language single-paradigm evaluation cannot surface. Applied to six watermarking schemes,
three open-weight generators, eleven languages spanning four scripts and eight
typological families, and both base and instruction-tuned regimes, the framework reveals failure modes that single-language single-paradigm evaluation cannot surface. Across detection and quality, observed disparity is predominantly between-family
on the typological partition, indicating that cross-lingual fairness gaps in watermarking are structural to language properties rather than idiosyncratic to particular
languages.Where the paper says it
Where the paper says it
Detection and Calibration Failures
The analysis of detection revealed that calibration failures can be distinguished from detection failures using a threshold-independent companion measurement. For instance, the AUC companion separates a calibration failure (TPR≈0, AUC≫0. p. 3 To diagnose this we report a
threshold-independent companion,
AUCc = PrSs(x
+) > Ss(x
−)
, x+ ∈X +
c
, x− ∈X −
c
,
estimated by the Mann-Whitney U statistic with average-rank tie handling. The TPR-AUC gap
separates a calibration failure (TPR≈0, AUC≫0. The TPR-AUC gap
separates a calibration failure (TPR≈0, AUC≫0. 5) from a detection failure (TPR≈0, AUC≈
0. Instruction-tuning collapses detection. The instruct-AYA panel shows mean TPR below 0. 7 in 16
of 18 cells and DI<. 8 ≥ 8 in 12 of 18, with non-SynthID cells dropping by roughly 0. 8 ≥ 8 in 12 of 18, with non-SynthID cells dropping by roughly 0. 78 in
mean TPR relative to their base counterparts.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Quality Preservation Dynamics
Quality preservation is measured across three disjoint paradigms: distributional (MAUVE), paired-semantic (BERTScore), and reference-perplexity. Distortion-free schemes like EXPEdit and SynthID are empirically the most distorting under every paradigm. p. 5 Distortion-free schemes are the most distorting empirically. EXPEdit and SynthID (the two
distortion-free schemes) sit at the bottom of the quality ranking under every paradigm in the base
regime. ,
SynthID-Gemma instruct, TPR 0. 602) show near-perfect quality preservation, consistent with the
watermark having less effect on the generation distribution.Where the paper says it
Where the paper says it
Structural Disparity and Robustness
The generalized-entropy decomposition of disparity over a typological partition reveals that the between-family share quantifies whether disparity is structural to typology (script, morphology, tokenizer interactions) or idiosyncratic. For cells with GE2 above the 10−3 floor, the between-family share sits between 61. p. 4 Structural rather than idiosyncratic disparity. For cells with GE2 above the 10−3 floor, the
between-family share sits between 61. For cells with GE2 above the 10−3 floor, the
between-family share sits between 61. 9% with median above 90% across both regimes. average and disparately across languages. Second, the typological partition captures the dominant
share of cross-language disparity across the grid: between-family ≥ 70% in 32 of 36 (scheme,
paradigm, regime) cells, with median≈ 85%, supporting the structural-disparity claim from §5. 01 (100 at α = 0. 05); paired-set quality measurements require at least
100 samples per side.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Societal Implications
The findings suggest that cross-lingual fairness gaps in watermarking are structural to language properties rather than idiosyncratic to particular languages. This means the same scheme offers materially weaker protection to speakers of some language families than others, and quality disparities mean the cost of being watermarked is unequally borne. p. 8 Societal impacts: Watermarking is a scalable mitigation for AI-generated text currently being
deployed, much of which is multilingual. A detection gap structural to typological family rather
than idiosyncratic to a single language means the same scheme offers materially weaker protection
to speakers of some language families than others, and quality disparities mean the cost of being
watermarked is unequally borne.Where the paper says it
Conclusion
In this work, the authors introduce the first systematic cross-lingual fairness audit of LLM text watermarking: a four-component evaluation framework comprising empirical FPR calibration, a threshold-independent AUC companion, three disjoint quality paradigms, and a generalized-entropy decomposition over a typological partition. The audit surfaces failure modes that single-language single-paradigm evaluation cannot surface. p. 9 eleven languages across four scripts and eight typological families, under two generation regimes. The audit surfaces failure modes that single-language, single-paradigm evaluation cannot, including
calibration failures misread as detection failures, distortion-free schemes that are empirically the most
distorting, cross-paradigm rank inversion, and disparity dominated by between-family variation.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Calibrate detection thresholds empirically per deployment context by using quantile 1-alpha of scores from unwatermarked text, denoted as tau c(alpha), to model deployment with a separate detector per language or a global threshold.
-
Implement a threshold independent companion measurement, AUCc = PrSs(x+) > Ss(x-), estimated by the Mann-Whitney U statistic, to distinguish calibration failures (TPR≈0, AUC≫0.5) from detection failures (TPR≈0, AUC≈0.5).
-
Utilize three disjoint quality paradigms—distributional (MAUVE), paired-semantic (BERTScore), and reference-perplexity—to measure text preservation, as no single metric can adjudicate quality across different distortion mechanisms.
-
Apply a generalized entropy decomposition of cross-language disparity over a pre-registered typological partition to reveal whether cross-lingual fairness gaps are structural to language properties rather than idiosyncratic.
-
Design systems where detection and quality axes are balanced by analyzing the regime spread, noting that
Base-regime spread is dominated by detection-side calibration failure
andInstruction tuning swaps the ranking.
-
Ensure that distortion-free schemes maintain substantially better quality preservation under finite sample distributional, paired-semantic, or reference-perplexity measurements compared to logit bias schemes.
-
Develop models capable of identifying which structural variable—tokenizer fertility, unwatermarked-generation surprisal under XGLM-7.5B, Latin script indicator, or language adherence rate—drives detection TPR and quality preservation on a per-language basis.
-
Design instruction tuning regimes to potentially compress the watermarked and unwatermarked score distributions together but not erase their separation, as
the instruct collapse is therefore predominantly calibrational.
Abstract
Watermarking schemes for large language model output are evaluated almost exclusively on English text using each scheme's detection threshold and a narrow set of quality measurements. Multilingual deployment exposes evaluation-design choices that are inconsequential on English but determine conclusions cross-lingually. We propose an evaluation framework with four components: detection thresholds calibrated empirically per deployment context, a threshold-independent companion measurement that distinguishes calibration failures from detection failures, three disjoint quality measurement paradigms (distributional, paired-semantic, and reference-perplexity), and a generalized-entropy decomposition of cross-language disparity over a typological family partition. Applied to six watermarking schemes, three open-weight generators, eleven languages spanning four scripts and eight typological families, and both base and instruction-tuned regimes, the framework reveals failure modes that single-language, single-paradigm evaluation cannot surface. Against a shape-matched null, detection disparity is predominantly between-family on the typological partition, indicating that detection gaps are structural to language properties rather than idiosyncratic to particular languages; quality disparity is large but not reliably family-structured.
Sources
- Robust Distortion-free Watermarks for Language Models
- Watermark under Fire: A Robustness Evaluation of LLM Watermarking
- Watermarking Without Standards Is Not AI Governance
- Who Gets Flagged? The Pluralistic Evaluation Gap in AI Content Watermarking
- A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models
- Qwen2.5 Technical Report
- BERTScore: Evaluating Text Generation with BERT
- Provable Robust Watermarking for AI-Generated Text
Related papers
- ALARM: Audio-Language Alignment for Reasoning Models
- ThinkSum: Probabilistic reasoning over sets using large language models
- Coherence boosting: When your pretrained language model is not paying enough attention
- From Facts to Insights: A Persona-Driven Dual Memory Framework and Dataset for Role-Playing Agents
- Improving General Role-Playing Agents via Psychology-Grounded Reasoning and Role-Aware Policy Optimization
- The Art of Asking: Multilingual Prompt Optimization for Synthetic Data