Microsoft Research

Ambiguous object selection in text-guided image editing is solved by directly annotating the canvas, achieving state-of-the-art performance in complex edits.

Append-only memory guarantees attention collapse in long-horizon video reasoning, but actively rewriting a bounded working memory from an external filesystem boosts hard-subset task accuracy by over 20%.

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.

Cracking LLM ideation requires more than high sampling temperatures: reinforcement learning over cognitive creativity axes expands scientific exploration diversity by 28% while boosting proposal originality by 66 percentage points.

Several open research questions are identified, including how to efficiently record incidents and how to determine whether vulnerabilities and incidents generalize, and privacy requirements are summarized and research directions for the secure and trustworthy deployment of AI agents are outlined.

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.

Naive agent-generated test feedback degrades SWE-bench performance by reinforcing shared hallucinations, but decoupling test generation from repair through role-specific RL converts a 3.9-point loss into an 11.4-point gain on open-weights models.

AutoSaddler achieves up to 10% performance improvement in LLM agents by automatically optimizing harnesses based on execution failure signals.

Achieving an 86.17% success rate in reproduction test generation, DPIAgent reveals that structured task separation can dramatically enhance performance in automated software engineering.

Achieving a single successful task completion in stateful workflows doesn't guarantee reliability, with many agents failing to maintain consistent performance across multiple attempts.

Harnessed agentic RL can boost agent performance significantly, as shown by a 14.6-point improvement in coding tasks with minimal training data.

Action policies in multilingual tool-using agents are more consistent than previously thought, but only for models above a certain size threshold.

Adaptive execution horizons can boost task success rates by over 11% while reducing runtime overhead in Vision-Language-Action models.

AtumAI slashes the onboarding time for datacenter control-plane policies from months to mere minutes while consistently outpacing expert-engineered solutions.

Evaluating AI agents in a dynamic enterprise environment reveals that static snapshots miss critical context, but a new system allows for real-time, persona-driven assessments of agent performance across time.

Training agents in deep, evolving environments can dramatically enhance their performance, with a 9B model achieving a 30.6% accuracy increase through targeted design.

Change2Task recovers 29.2% more verified coding tasks than traditional methods, streamlining the training of coding agents.

The rise of LLM agents threatens to erase the clarity of authorship and accountability in collaborative knowledge work, raising urgent questions about intellectual integrity.

Specula uncovers deep bugs in system code that traditional methods often miss, revolutionizing formal specification generation.

OpenForgeRL allows researchers to train AI agents in real-world environments with unprecedented ease and efficiency, revealing that some harnesses are significantly harder to learn than others.

PRTA outperforms traditional and LLM-based recommendation systems by effectively leveraging multiple models through a central LLM planner, enhancing personalization without the pitfalls of hallucination.

Strong performance in static evaluations masks a critical flaw: LLMs struggle to adapt to evolving user intent during multi-turn interactions.

Real-world deployment of LLM-based agents reveals critical safety and reliability challenges that traditional benchmarks overlook.

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

TRACE transforms how long-horizon agents are trained, leading to a remarkable performance increase on complex tasks without the need for supervised fine-tuning.

STRACE transforms noisy execution traces into high-fidelity optimization contexts, yielding a 1.4x improvement in agent success rates on complex tasks.

ReOPD turns the costly process of agent-environment interaction into a reusable offline resource, achieving faster and more efficient multi-turn distillation.

Despite high diagnostic accuracy, LLMs fail to choose valid recovery actions for over 60% of incidents, exposing a critical flaw in their operational utility.

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

Despite the promise of advanced AI agents, even the best performers struggle, with Codex GPT-5.5 only achieving a 42% success rate in complex healthcare tasks.

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

Bridging the Context Gap in T2I models, Qwen-Image-Agent achieves state-of-the-art performance by intelligently constructing context from user input and external sources.

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

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

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

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.

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

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.

Current models struggle with hybrid interface tasks, achieving only a 41.2% success rate, underscoring a critical gap in CUA evaluation.

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.

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

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.

OpenWebRL-4B sets a new benchmark for open-source visual web agents, achieving impressive success rates with minimal initial data while outperforming larger-scale competitors.

SeClaw reveals that existing benchmarks fall short in capturing the complexities of agent behavior, enabling a more nuanced evaluation of security risks in autonomous systems.

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.