Microsoft Research

Even state-of-the-art LMMs struggle with reasoning-informed visual editing, with the best model only achieving 56.6% accuracy on a new comprehensive benchmark.

Choice-based models like Jev can accurately identify graph adjacencies but fail to grasp broader structural correctness, even with joint graph-text inputs.

Standard next-token prediction on an agent's self-generated explanations outperforms GRPO on SWE-bench in half the training updates—even bootstrapping on tasks where every initial rollout failed.

The latency with respect to the resource dynamics of processing multiple requests and tasks concurrently is characterized and new optimization opportunities that exploit the resource dynamics of tasks are demonstrated: CPU-aware tool admission and task-aware CPU allocation.

Registers are implemented as dedicated fixed-position tokens whose continuous hidden states are trained to carry reasoning progress across generation chunks, which are especially effective for bounded code generation, where correct programs usually span several chunks.

This work introduces MindTopo, a benchmark of topological intuition across five properties grounded in cognitive science and formal topology: continuity, separation, order, enclosure, and knots, which performs better on reasoning than on planning, and the best-performing model remains far below observed human performance.

Credit-addressable reasoning boosts multimodal geometry accuracy by over 8 points, revealing the critical role of structured learning in complex tasks.

Abstaining from uncertain predictions can enhance LLM accuracy in medical applications by 9.6 percentage points, transforming uncertainty into a strategic advantage.

Self-evolving rubric rewards can dramatically enhance audio reasoning in models, outperforming traditional methods by adapting to the model's evolving capabilities.

Agentic reasoning tools struggle with complex financial documents, revealing substantial gaps in their capabilities that could impact decision-making in finance.

GFlowRL achieves unprecedented stability and performance in large language models by eliminating the problematic learned partition function, setting a new standard for GFlowNet-style reinforcement learning.

Data synergy can either amplify or diminish model performance, revealing that the right dataset combinations are crucial for optimal language model training.

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

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

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.

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

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

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

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

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.

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

AI-driven scientific discovery is closer than you think, but current systems still struggle with reproducibility, cross-domain robustness, and accountable scientific closure.

Don't let your SWE agent drown in context: SWE-AGILE maintains performance on multi-turn software engineering tasks by dynamically managing reasoning context with a novel sliding window and compressed reasoning digests.

LLMs still fail to grasp research-level mathematics, with top models scoring below random chance when superficial pattern matching is removed, even with access to proof sketches.

Particle filtering reveals a fundamental limit to inference-time sampling methods for LLMs, suggesting that simply increasing the number of samples has diminishing returns.

GUI agents can achieve significantly stronger task-solving capabilities through carefully designed post-training and data curation, without relying on costly online data collection.

AgentOS reimagines LLMs as reasoning kernels within a structured OS, offering a blueprint for more robust and scalable AI agents.

LLMs can't reliably debug code in long contexts (64k-128k tokens) even with perfect information retrieval, despite impressive performance in agentic workflows that decompose the task.

Scaling laws hold for interest modeling: bigger LLMs and more inference-time sampling consistently boost news recommendation quality, and can be distilled into smaller, deployable models.

Ditch the army of task-specific models: AdNanny shows a single, reasoning-centric LLM can handle diverse offline advertising tasks with improved accuracy and reduced manual effort.