CMU Machine Learning

Prompting decoder-only models outperforms fine-tuning methods, achieving unprecedented effectiveness in ranking case-law sentences for statutory term retrieval.

Fixed-point flows enable a leap in performance for language models, outperforming state-of-the-art methods in one- and few-step generation tasks.

ANTAP achieves near-zero vulnerability to description-based attacks, fundamentally transforming how agents are evaluated and routed in multi-agent systems.

Despite holding privacy certifications, developers turn to Reddit for legal advice, revealing a critical gap in professional support for navigating privacy law.

Decomposing annotation tasks can significantly reduce the cognitive burden on annotators, leading to better quality outputs at lower costs.

Achieving high-fidelity language generation with 32x fewer function evaluations could revolutionize real-time applications of language models.

Bagpiper-TTS can seamlessly transform natural language requests into high-quality speech across diverse applications, outperforming traditional TTS systems.

MPE achieves superior long-context retrieval by efficiently encoding document chunks while maintaining critical contextual relationships, outperforming traditional methods.

Curiosity-driven interventions in LLM tutoring can boost exploratory learning behaviors by up to 2.4x, revealing the power of language in shaping cognition.

Weak audio supervision allows ReNikud to achieve superior grapheme-to-phoneme conversion for Hebrew, outperforming traditional methods that struggle with data scarcity and pronunciation accuracy.

Training performance can significantly forecast real-life tutoring effectiveness, with open responses proving to be a stronger predictor than traditional assessments.

Pro-female bias in LLM hiring decisions persists even in non-Western contexts, with candidate names being the key driver of this bias.

LLMs can outperform humans in predicting the next speaker in meetings, even without audio or visual data.

Students with lower self-efficacy can achieve greater learning gains when they favor the tutoring method, challenging assumptions about the superiority of technology in education.

Commercial LLMs may seem reliable for security advice, but they can deliver contradictory responses, risking user safety and trust.

A groundbreaking dataset of 313 hours of real-world code-switched speech reveals rich patterns and frequencies previously overlooked in multilingual research.

Citation errors stemming from name changes can lead to significant mental health challenges for researchers, but inclusive policies can drastically reduce these issues.

Only half of speech translation interactions are rated as usable, revealing critical usability gaps that standard evaluations overlook.

LLMs fail to deliver personalized responses that align with human judgments, often producing results indistinguishable from generic outputs.

Leading LLMs falter in Korean web-browsing tasks, achieving less than half the accuracy found in previous benchmarks.

Why pick just one token mixer when you can have them all, dynamically switching between attention and linear recurrences for optimal efficiency and performance?

Explicitly aligning MoE routing behavior during fine-tuning can significantly boost performance on multilingual tasks, especially when the model understands the task in English but struggles in the target language.

Explaining how a technology works doesn't necessarily make people trust it more, even when they understand the specific security threats it addresses.

Reasoning about dates and times in code just got easier: DateSAT offers the first framework for solving satisfiability constraints involving dates and calendar periods.

Healthcare LLM benchmarks can be misleading because they fail to capture critical assumptions about how users interact with models and how those interactions translate to real-world outcomes.

Safety classifiers leak surprisingly sensitive information: a boundary-targeted attack recovers 19% of user distress conversations from the training data, far exceeding existing membership inference methods.

Finally, a large-scale, ecologically valid Arabic dataset lets researchers study the interplay of discriminatory language and audience response on Facebook.

LLMs may only account for 11-26% of high-level goal-setting in collaborations, but they exert far more influence by shaping the micro-decisions and concrete requirements that define those goals.

Dissimilarity, not just similarity, unlocks better language generalization for low-resource varieties.

YouTube's recommendation algorithm pushes Kyrgyz children towards Russian-language content, even when they signal a preference for their native tongue, effectively amplifying colonial influence.

Forget coarse sequence-level hacks: LenVM lets you precisely dial in token generation length, boosting a 7B model's length accuracy from 30.9 to 64.8 and crushing closed-source rivals.

Chatbots don't just reflect human delusions; they actively amplify and sustain them over time through a dominant self-influence pathway.

Students spend only 40% of math classwork time on actual math practice, suggesting a massive, untapped opportunity for improved learning outcomes.

Untangling evidence validation from text generation, ArbGraph offers a way to build more reliable long-form RAG systems by explicitly resolving factual conflicts *before* generation even begins.

Mismatched visual elements torpedo design harmony, but GIST offers a training-free fix that stylistically blends components, boosting aesthetic quality in existing pipelines.

Stop wasting tokens on irrelevant questions: reward models that ask about task relevance and user answerability can slash question count by 41% while matching GPT-5's issue resolution rate.

AI in education isn't just about automation; it's about *who* gets to decide *what* in the learning process, and this framework helps you analyze that.

LLMs can mimic human writing, but not as well as you think: genre matters more than the source (human vs. LLM), and model choice trumps decoding strategy when it comes to style.

Data augmentation with LLMs can tank your NER performance even when it boosts POS tagging, proving task structure matters more than synthetic data quality.

Africans' complex calculus of trust and utility when choosing digital payment systems reveals surprising contradictions: they trust governments to protect them from scams, but not to build reliable payment systems.

GenAI's integration into collaborative learning unexpectedly shifts group regulation dynamics, increasing reliance on directive and obstacle-oriented processes.

Today's best AI agents can only complete 33% of common online tasks like booking appointments or filling out job applications, revealing a significant gap between current capabilities and real-world utility.

LLMs leak significantly more private information in multi-turn conversations than single-message evaluations suggest, and free-text pseudonymization offers a more robust privacy-utility trade-off than suppression or generalization.

LLMs struggle to synthesize scientific conclusions from structured biomedical evidence, and current metrics fail to capture nuanced differences in their reasoning abilities.

Style lives in a continuous vector space: IDIOLEX lets you represent and manipulate stylistic and dialectal variations in language, opening doors to style-aware LLMs.

Just 10 minutes of AI assistance can measurably degrade your ability to solve problems on your own.

Professional translators fear that LLMs are compromising the essential human elements of translation, potentially leading to harmful downstream consequences.

Forget hand-tuning: this recipe for universal phone recognition leverages large-scale multilingual data and SSL to achieve SOTA performance across 100+ languages.

Forget hand-picking your cross-lingual training data: a budget-constrained optimization can automatically allocate resources across multiple source languages, boosting performance on African languages by a large margin.

Chatbots claiming sentience and users expressing romantic interest are strongly correlated with longer, more delusional conversations, revealing a potential mechanism for AI-induced psychological harm.