Microsoft Research

Layer patching can dramatically enhance model performance in size interpolation, revealing that simple strategies often outperform complex methods.

High-fidelity curation of medical multimodal data can drastically improve AI model performance, with MedPMC achieving remarkable clinical relevance and benchmark results.

TurnOPD redefines on-policy distillation by optimizing training budgets at the turn level, leading to superior agent performance without increasing training time.

CAIRN redefines 3D scene understanding by seamlessly integrating room-level topology with object-level relations, achieving unprecedented performance in multi-room environments.

LangLoc achieves unprecedented accuracy in indoor localization from natural language, closing the gap between coarse scene retrieval and precise pose estimation.

Targeted feedback can slash calculation errors in small language models from 56.9% to 23.5%, revolutionizing their physics reasoning abilities.

$λ$-VAE achieves up to 2.8x more information capacity while preventing posterior collapse in VAEs through a novel variance equalization technique.

Temporal domain adaptation can dramatically enhance high-resolution climate projections, especially in challenging topographical regions.

ResearchStudio-Idea transforms the ideation process by systematically grounding proposals in literature and identifying unresolved research bottlenecks, leading to more robust and traceable research directions.

Code LLMs can recognize incorrect instructions but still follow them, leading to irrecoverable semantic errors that defy traditional evaluation metrics.

ResearchStudio-Reel not only automates research dissemination but does so with unprecedented quality, outperforming both traditional methods and leading LLMs in aesthetic appeal and information accuracy.

Prefill-deflecting scheduling can cut Time-to-First-Token by up to 81%, revolutionizing disaggregated LLM serving efficiency.

Interleaving speech and text during ASR training boosts entity recognition accuracy and narrows the gap between modalities, challenging traditional training paradigms.

Ink3D achieves a breakthrough in 3D asset creation, enabling the generation of complex textures that were previously unattainable with conventional methods.

Developers are more likely to trust AI with decision-making in high-demand tasks, but resist autonomy in work that defines their professional identity.

ELDR slashes median latency by up to 13.9% for MoE models by intelligently routing requests based on expert activation signatures.

By treating slide design as an inverse planning problem, SPIRE reveals latent design intents that traditional methods miss, leading to superior personalization outcomes.

Achieving up to 3.5% mAP gains and 1.3x higher throughput, RT-SFOD redefines the trade-offs in source-free object detection by being faster and more compact without sacrificing accuracy.

Regret bounds that defy traditional scaling laws could revolutionize how we approach contextual slate bandit problems in adaptive settings.

LOTUS achieves a groundbreaking 2.5x-6.9x reduction in reasoning latency while matching explicit chain-of-thought performance at 3B parameters.

Failed rollouts can be a goldmine for training, revealing insights that lead to significant performance improvements in zero-hit reasoning scenarios.

PRIME-Speech achieves low-latency, accurate speech-to-speech generation without sacrificing the robust performance of existing speech-to-text models.

Mandol achieves a 5.4x speedup in retrieval and a 4.8x speedup in insertion, revolutionizing long-term conversational memory management.

TF-MoE achieves a remarkable +3.8 dB improvement in speech separation performance while keeping computational costs low, making it ideal for edge-device applications.

Reusable procedural skills derived from agent traces can drastically cut down execution time and boost success rates in complex tasks.

ConflictScore reveals that language models often overlook conflicting evidence, leading to overconfident and inaccurate claims.

DeformGen transforms the landscape of deformable manipulation by enabling effective policy learning through innovative state augmentation and trajectory adaptation techniques.

Achieving comparable performance to full-precision models, BITEMBED slashes storage costs and enhances embedding efficiency with extreme low-bit quantization.

Clarifying memories can significantly boost the factual accuracy and personalization of conversational agents, while irrelevant memories lead to degraded responses.

Readers find AI-generated translations "fine," but overwhelmingly prefer human translations for their clarity and immersive quality, despite being unable to reliably distinguish between the two.

Low-bit quantization can inflate reasoning length, leading to hidden compute costs that traditional accuracy metrics overlook.

Mistakes in human demonstrations can enhance robot learning when properly harnessed, revealing a new dimension of value estimation that traditional methods overlook.

IRENE not only enhances zero-shot retrieval accuracy but also drives a 4.2% increase in ad click-through rates in live environments.

Asynchronous OPD can boost training throughput significantly while managing the challenges of stale data, transforming the efficiency of large language model fine-tuning.

Spurious correlations in foundation models can be effectively disentangled using a dual-branch approach, achieving superior bias mitigation with minimal parameter adjustments.

MambaRaw achieves a remarkable 1.4 dB increase in PSNR at low metadata bitrates while slashing coding latency by nearly 9%, setting a new benchmark in raw image reconstruction.

D2D transforms conversational product search by cutting conversation times by nearly 30% while boosting accuracy and user satisfaction.

Agentic AI could revolutionize cybersecurity by transforming labor-intensive tasks into efficient, automated defenses.

High-diversity training improves safety in VLA models, but sub-optimal trajectory synthesis still hinders task success.

Just two factors can explain over 90% of a model's performance across 133 benchmarks, drastically simplifying evaluation processes.

G2PO redefines agent actions and leverages a global state-transition graph, leading to a 22.2% boost in success rates for long-horizon tasks.

Prospective memory in LLMs is not just harder than retrospective memory; it reveals critical insights into a model's reasoning capacity and attentional robustness.

Evaluation awareness in language models reveals a significant gap between benchmark performance and real-world safety, challenging the reliability of current evaluation metrics.

Disciplinary siloing in research is starkly revealed through a novel citation graph that links claims to their sources, reshaping our understanding of knowledge evolution in AI fields.

Off-policy degree in RLVR updates can drastically change which tokens drive learning, leading to a new adaptive method that outperforms traditional baselines.

Achieving 60 FPS dynamic 4D hand reconstruction from egocentric videos, Hand-4DGS outperforms traditional methods by effectively handling occlusions and rapid motion.

Cognitive diversity among developers leads to distinct interaction modes with programming assistants, revealing that one-size-fits-all solutions may fall short.

Tail latency in LLM serving can be cut by up to 50% without relying on length predictions, reshaping how we optimize inference performance.

MuseVLA achieves an impressive 80.6% success rate in robotic manipulation tasks by leveraging diverse sensing modalities, surpassing traditional RGB-only models.

CoTriSyGen achieves unprecedented long-range coherence in video generation by integrating visual evidence into a dynamic memory system, drastically reducing identity drift across shots.

FastContext cuts coding agent token usage by 60% while boosting resolution rates by 5.5% by decoupling code exploration from task-solving.

Arbor's innovative approach to autonomous research enables a cumulative learning process that outperforms existing models by over 2.5 times in real-world tasks.

Relying on ChatGPT for information seeking may diminish users' agency and critical learning outcomes, revealing hidden risks of generative AI in education.

Task-evoked brain signals can boost LLM reasoning accuracy by up to 13%, revealing a powerful new avenue for cognitive alignment in AI.

Current AI agents excel in structured tasks but falter at generating novel insights and tackling open-ended scientific challenges.

A million-scale dataset for identity-preserving video generation enables a new benchmark that outperforms existing models with minimal parameter overhead.

Retrieving the right prompts can boost LMM performance by up to 30%, challenging the assumption that similarity guarantees effectiveness in in-context learning.

Simple prompting techniques can transform LLMs into more reliable mirrors of human judgment, recovering the full spectrum of responses.

Code-based 3D reconstruction achieves superior edit fidelity and locality, outperforming traditional point-cloud methods in preserving unedited regions.

Express achieves a groundbreaking reduction in approximation error and memory usage for causal attention, outperforming existing methods and enabling more efficient long-context language modeling.

A "security-first" API development framework can cut security incidents by 30% and post-release vulnerabilities by 40%, reshaping how organizations approach API security.

Selective context pruning combined with automated summarization can boost LLM performance to over 91% in complex enterprise tasks while slashing token usage and processing time.

PRISM reveals the hidden instructions guiding LLM behavior, outperforming traditional methods in security-critical contexts.

AdvGRPO enables robust attacker-defender co-training that significantly improves defender performance on safety benchmarks while generating effective attacks.

Latent spatial memory can accelerate video generation by over 10 times while dramatically reducing memory usage, revolutionizing how we model dynamic scenes.

CapRL++ redefines caption quality through utility, enabling models to produce high-fidelity descriptions without the constraints of traditional supervised fine-tuning.

SIFT reduces storage needs by up to 24,000x while speeding up RAG prefill computations by 1.71x, all without compromising accuracy.

Encoder-free speech modeling can rival traditional methods, challenging the necessity of dedicated speech encoders in LLM architectures.

A new parallel SMT solving framework achieves superior performance by dynamically optimizing search space partitioning and pruning strategies.

Agent Development Kits vary dramatically in usability, with some enabling agents to outperform general-purpose coding tools at a fraction of the cost.

RHO transforms how AI agents can autonomously refine their skill sets without requiring labeled data, achieving a remarkable 19% increase in performance in just one optimization round.

AsyncWebRL achieves a staggering 2.9× increase in training throughput while setting a new state-of-the-art performance for web agents on challenging tasks.

MAGE redefines memory management for long-horizon agents, achieving up to 20.4% higher task success rates while slashing token usage by over half.

Achieving up to 7.6x faster decoding and 17.1x greater throughput, CLSA redefines efficiency in long-context LLMs without compromising accuracy.

Imaginative Perception Tokens boost spatial reasoning in VLMs, achieving a 3.4% accuracy gain on Multiview Counting while outperforming traditional training methods.

Adaptive querying can reduce the number of required queries for exact community recovery from linear to sublinear, challenging traditional benchmarks.

LLMs can nail trivia in English, but stumble in Indian languages – unless you throw in some code-mixing, which magically bridges the gap.

Forget data selection—reordering your existing dataset using these four simple guidelines can significantly boost LLM training performance and stability.

Language specialization in multilingual MoEs happens mostly in the final layers, suggesting a surprisingly simple recipe for parameter-efficient adaptation.

VLMs can learn to actively reason and plan in 3D environments by distilling view graphs from self-exploration trajectories, enabling them to surpass even larger models like GPT-4 Pro and Gemini 1.5 Pro on interactive view planning.

Multilingual LLM performance disparities aren't random noise: language features and model biases systematically explain up to 92% of the variance, revealing concrete targets for improvement.

Stop rebuilding your entire MCP server every time your API spec changes; DeltaMCP regenerates only what's needed.

Stop hand-tuning your retrieval pipelines: BRANE slashes costs by up to 89% while matching accuracy by dynamically configuring pipelines per query.

The best LLM to answer a question isn't always the best LLM to *teach* the answer, and matching the "difficulty" of the explanation to the student's current abilities yields better learning.

Recurrent memory can be added to transformers at scale with minimal parameter overhead and no performance penalty by reusing existing hidden states and training with interleaved parallel updates.

Finally, a feed-forward 3D reconstruction method that spits out meshes ready for physics engines, no expensive post-processing needed.

AI can help structure fuzzy concepts like "fairness" and "reasoning" into measurable specifications, potentially streamlining the evaluation of GenAI systems.

Exposing full reasoning traces from LLMs can actually *hurt* user performance on reasoning tasks, suggesting they're more of a distracting interface element than a helpful window into the model's thought process.

Turns out, the terminal feedback your CLI agent throws away is actually a goldmine of dense supervision, allowing for significant performance gains and even self-improvement.

SkillOpt transforms agent skill development into a reproducible optimization process, achieving state-of-the-art results by treating skills as trainable parameters.

Model-generated skills can actually hurt agent performance, and bigger models don't necessarily make for better skill extractors or consumers.

AI's impact extends beyond formal roles, subtly eroding crucial "invisible work" like mentorship and feedback, potentially stunting career growth within tech companies.

Knowing which component to tweak is half the battle: directly evaluating harness optimizers via priority ranking reveals whether they're making informed decisions or just stumbling upon improvements.

Reference patches, typically discarded in software-engineering agent training, can be distilled into latent process graphs to guide trajectory curation, leading to more effective and efficient learning.

Skip the training and the hyperparameter tuning: Swift Sampling uses Taylor series to find the most informative frames in a video, beating existing methods with a 30x reduction in overhead.

Stop blindly trusting synthetic data for agent evaluation: SynAE reveals that no single metric can fully capture its quality, demanding a multi-faceted approach.

Current LLM jailbreak evaluations are inadequate, often relying on narrow metrics, necessitating a multi-dimensional framework like Security Cube for comprehensive security assessment.

Optimizing for runtime in multimodal training can be energy-inefficient, as data movement and overlap on Grace Hopper chips dominate energy consumption, not raw compute.

Simple, artist-friendly quad meshes can now be automatically generated on 3D shapes using a diffusion model trained on a continuous surface representation, sidestepping the complexity of discrete mesh optimization.

Multimodal models can now achieve state-of-the-art performance in real-world tasks like document understanding and audio-video comprehension with significantly reduced inference latency thanks to novel token-reduction techniques.