CMU Machine Learning

Ego-centric video scaling saturates agent modeling long before object interaction fidelity, revealing a fundamental bottleneck in current world model training paradigms.

Microscopic robots can now execute complex tasks from natural language instructions with unprecedented accuracy, achieving 6.30-pixel RMSE using zero new labels and outperforming specialized models by over 50%.

Activations, not just descriptions, unlock the prediction of how LLM alignment generalizes to unseen values, revealing a structured value space.

Pretraining neural CFD surrogates across diverse domains unlocks substantial gains in zero- and few-shot generalization, drastically reducing data requirements for new applications.

Agent experience, even when uniformly unrewarded, can be distilled into predictive foresight to dramatically improve learning in sparse reward settings.

Even clonable quantum proofs can outperform classical ones against a classical oracle, revealing a subtler source of quantum advantage beyond mere unclonability.

Single-photon cameras transition from ambiguous to near-deterministic scene recovery as photon budget increases, a phenomenon precisely modeled by a novel Hypergeometric generative framework.

Multimodal agents can achieve >90% reasoning latency reduction and comparable task performance by selecting from pre-defined reasoning steps instead of generating them freely.

Robots can now reliably navigate through buildings with glass, avoiding collisions and phantom obstacles by accurately mapping transparent surfaces.

Autonomous agents can recursively self-improve LLM-generated games, achieving superhuman performance and drastically cutting generation costs by internalizing development experience.

Faster computer-use agents are bottlenecked by their speed, but existing benchmarks are irreproducible; this new standardized framework reveals counterintuitive speed-efficiency trade-offs and that current open-weight models lag behind.

State-of-the-art VLMs struggle to execute scientific research workflows, revealing a critical gap in their ability to interact with specialized software and produce verifiable results.

A single visual interface, action views, bridges the gap between video generation models and diverse robotic embodiments, enabling unified action modeling and execution.

Multi-action weakly-coupled MDPs can shatter the theoretical $O(1/\sqrt{N})$ performance barrier to reach an $O(1/N)$ optimality gap simply by ditching heuristic priority ordering in favor of locally linear mean-field dynamics.

Native 8-bit LLM training silently destabilizes at billion-parameter scale due to softmax gradient drift, but restoring the zero-row-sum invariant completely eliminates the FP8 attention gap without architecture modifications.

Standard BF16 FlashAttention silently breaks a fundamental zero-sum symmetry in the softmax backward pass, corrupting query gradients by over 200% and triggering thousandfold gradient blowups late in pretraining.

Contemporary frontier speech models still catastrophically fail on living Arabic dialects, with top systems like GPT-4o achieving an error rate no better than 35.0% WER across 17 regional vernaculars.

Autonomous model post-training can outperform human-crafted instruct baselines on hard coding benchmarks while completely eliminating reward hacking once experimental exploration is governed by a sandboxed OS and a dynamic peer-review DAG.

This first study of Web search across four major conversational platforms is presented, combining real-world user interactions with controlled experiments using the same platform's models by their APIs (invitro), finding that Web-search decisions vary substantially across platforms and models.

How chatbot misgendering takes place and how, in response, users engage in ongoing work to curate their gender presentation to prevent and repair misgendering is documented.

Experiments show that MoRE consistently achieves lower perplexity and stronger downstream performance than standard MoEs and state-of-the-art weight-sharing architectures at matched compute and parameter budgets, with only minimal modifications to existing MoE implementations.

This paper presents an end-to-end approach for generating context-specific large language model (LLM) benchmark datasets by combining expert input with synthetic data generation by introducing a schema eliciting key information about the goals, scope, and context of an evaluation task.

LACE (Layer-Adaptive Codec Encoding), a dynamic frame rate codec that applies an independent compression step at each quantization layer, enabling layer-specific segmentation boundaries, is proposed.

R-U-Net is proposed, an ECG delineation model that pairs a ResNet-18 encoder with a U-Net decoder and shows that decoder design contributes more to performance gains than the evaluated SSL methods, motivating further exploration of architectures for ECG delineation.

Safety monitors miss critical risks like sandbagging and data leaks not because LLMs lack capability, but because hyperproperty detection fundamentally demands an executed second trace and an explicit comparative procedure.

Two AI systems with nearly identical performance can have drastically different human oversight needs, revealing hidden costs in deployment readiness.

Free-form language reasoning can dramatically enhance robotic manipulation, outperforming traditional instruction-based methods in complex tasks.

HSR boosts robot manipulation success rates by over 21% by leveraging hierarchical skill retrieval, even with minimal task-specific data.

GCA reduces communication overhead in federated learning by up to 99.15% while simultaneously enhancing data protection and improving model accuracy.

By making environment design a learnable process, SPADE unlocks a new frontier in self-improvement for language agents, leading to substantial performance gains across diverse tasks.

A 5-step sampling schedule can deliver nearly the same quality as a 50-step one, slashing inference costs by 90%.

Open-vocabulary monocular 3D detectors mislabel correctly localized objects due to prompt sensitivity, revealing a critical gap in their semantic understanding.

Closing the loop on robot manipulation failures, VLCP rewrites control code in real-time, achieving a tenfold increase in success rates over traditional methods.

HarnessEval-W transforms world model evaluation from mere scoring to a transparent reasoning process that mirrors human judgment.

Achieving a new upper bound of $\omega < 2.371177 could redefine our understanding of matrix multiplication efficiency.

Current AI-generated videos that mislead viewers are also the most challenging for existing detection systems to identify, revealing a critical vulnerability in misinformation defenses.

Explicitly optimizing for target function smoothness can dramatically enhance the performance of models on tabular data.

Social Gym reveals that even top-performing LLMs struggle with consistent social reasoning across diverse multi-agent scenarios.

High speech overlap isn't the primary challenge in cocktail-party scenarios; innovative audio-visual strategies and large language models can cut recognition errors by 57%.

A 'severity-first' approach to risk assessment in the EU AI Act could expose vulnerabilities to manipulation by AI providers, risking regulatory effectiveness.

SJRL not only overcomes collision challenges in multi-agent navigation but also adapts dynamically to real-world constraints, outperforming traditional methods in complex environments.

Formal verification of a compiler for asynchronous dataflow could redefine the reliability and efficiency of parallel computing architectures.

Path-level pretraining in MultiPathFormer leads to a dramatic 59% improvement in wireless propagation estimations, reshaping the landscape of wireless foundation models.

Skill-switching accuracy in LLMs drops significantly on complex tasks, but a new training approach boosts performance from 34.4% to 68.4% on challenging benchmarks.

H2S achieves a remarkable 48.54 mAP in audio-visual instance segmentation, setting a new benchmark in the field.

The rise of AI in software development is rendering traditional software measurement assumptions obsolete, necessitating a radical rethink of how we validate our findings.

No randomized polynomial-time algorithm can overcome the condition-number barrier in sparse least-squares optimization, confirming a long-standing conjecture.

Retaining the right contextual information can boost long-context modeling performance by over 20% compared to traditional methods.

Agentic Commerce World reveals that process-level evidence is crucial for accurately evaluating AI agents in dynamic market environments, challenging traditional reliance on final outcomes alone.

Code writing outperforms all other active learning methods in programming education, leading to significantly better learning outcomes.

SRA can double the operational capacity of automated warehouses while slashing optimization time from hours to minutes.

Achieving high-quality reranking with a 30B MoE model is now feasible on an academic budget, outperforming traditional dense models in efficiency.

GenAI could either reinforce linguistic hierarchies or become a powerful tool for promoting diversity in academic writing, depending on its design and governance.

Coding agents may boost productivity, but they risk diminishing developers' understanding and long-term coding skills.

VLMs can fabricate highly biased medical diagnoses based solely on demographic descriptors, revealing alarming implications for clinical trustworthiness.

Downstream performance hinges on SSL pre-training language, not the NAC training language, enabling efficient cross-language model reuse.

Proprietary MLLMs may achieve high diagnostic accuracy, but they still struggle with reliable clinical reasoning, revealing significant gaps in their practical utility.

Achieving over 99% fulfillment of observation requests in decentralized satellite scheduling reveals a transformative approach to large-scale DCOPs that outshines traditional methods.

Achieving 19 times lower inference latency than diffusion models, UniGen-AR redefines the efficiency of unified visual generation.

Most developers edit AI-generated code within 15 minutes, often discarding the original completions entirely, highlighting a critical gap in LLM training data.

Retrieval harm can significantly impact performance, and simple baselines often outperform complex models in Active RAG systems.

Different views of the same problem reveal hidden reasoning paths, enabling VLMs to achieve unprecedented accuracy in multimodal reasoning tasks.

Adapters may store less than half the information expected, with their capacity heavily influenced by parameter placement rather than quantity.

Token generation timing can leak critical architectural and optimization details from language models, exposing vulnerabilities in their deployment.

A lightweight vision-language model can effectively identify UI principle violations, achieving an impressive F1 score of 84% across multiple quality dimensions.

Attention-supervised finetuning reveals that model architecture significantly impacts the plausibility of explanations in media bias detection, challenging assumptions about scale and performance.

μSTM achieves high performance without sacrificing safety or usability, allowing for general data types and deferred aborts in transactions.

Smoothing the loss landscape with chain-of-thought reasoning can significantly enhance LLMs' ability to navigate complex moral dilemmas.

A unified framework reveals that robust downgrading mechanisms can coexist with full-strength non-interference, transforming our approach to program security.

AniGS transforms static 3D reconstructions into dynamic, immersive environments by seamlessly integrating ambient motion without compromising structural fidelity.

The PanAf-SBR dataset reveals that fine-grained social behaviour recognition in wild great apes can be significantly improved through targeted cross-dataset pre-training.

EvolvingWorld reveals that an open-schema framework can drastically enhance the coherence and depth of character and world interactions in long-horizon literary simulations.

Nine out of ten AI-selected modeling changes in materials science remain effective when tested on unseen data, showcasing the potential for reusable AI-driven discoveries.

Visualizing navigation structures transformed practitioners' approach to accessibility, shifting their mindset from compliance to design innovation.

Missing treatment data can lead to biased policy estimates, but leveraging partially-observed units with MAR estimators offers a more efficient and valid approach.

LLMs struggle with historical analogy retrieval due to a focus on surface features, but the new CANA framework boosts their performance by 10% through causal understanding.

EMAGN reduces the complexity of traffic forecasting models from quadratic to linear, enabling larger configurations without sacrificing performance.

Requential coding can compress billion-parameter models to sizes orders of magnitude smaller than traditional methods, revealing the hidden structure in datasets.

Indirect data poisoning can enable scientific fraud at an unprecedented scale, with a staggering 49.56% success rate in corrupting AI-driven research outputs.

TALRanker autonomously navigates the trade-off between efficiency and accuracy in document reranking, achieving state-of-the-art results without the latency of excessive tool calls.

Phonological features can be extracted from self-supervised speech models in under a minute, achieving state-of-the-art performance in phone segmentation and recognition.

LightCrafter achieves superior video relighting by integrating PBR with diffusion models, enabling intricate lighting control and long-form temporal consistency without the need for extensive training data.

Agents can now work independently on data changes while humans maintain oversight, revolutionizing collaborative data management.

Every program in the new Calf framework must preserve both abstraction and potential, revolutionizing how we approach cost verification in type theory.