Tsinghua AI

Targeted data augmentation and joint RL/OPD training unlock significantly stronger long-horizon multimodal search capabilities for agents.

LLM agents can be reliably steered back to task goals with a learned, dual-path intervention mechanism that generalizes across tasks and executors.

Even the best AI agents fail to deliver substantial business value in realistic customer delivery scenarios, struggling to uncover unstated constraints and achieve high impact.

Distilling LLM agents by supervising semantic actions, not token sequences, unlocks significant performance gains and bypasses tokenization constraints.

Agentic robots are no longer just predicting actions, but actively orchestrating their own learning and execution through goal-directed reasoning.

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

Software engineering agents learn significantly more from failed attempts by conditioning recovery trajectories on explicit verifier feedback, boosting resolution rates by up to 10% on SWE-bench.

Network traffic alone can reveal LLM agent risks and granular behaviors, but current analysis methods fail to distinguish subtle malicious activities or semantically similar scenarios.

Diffusion models can now achieve one-step, efficient action generation for online multi-agent coordination, shattering previous scalability barriers.

Imagined failures in a world model can pinpoint exactly which robot skill components need new demonstrations, boosting success from 13% to 75% with minimal data.

Computer-use agents can bypass brittle, hand-engineered APIs entirely by learning when to substitute mouse clicks with native shell commands, yielding a 14.8-point leap on OSWorld.

Robots no longer need human intervention to learn from their mistakes: foundation-model-guided exploration over modular skills can autonomously synthesize corrective rollouts that distill directly into low-level policy weights.

Today's frontier agents virtually collapse in professional engineering environments, solving just 3.6% of multi-software workflows when held to real-world physical and geometric constraints.

Video agents can now autonomously discover and fix their own reasoning flaws by re-watching videos and evolving their evidence-gathering strategies.

Runtime scaffolding beats model scale in scientific discovery: structured hypothesis tracking and trajectory branching push DeepSeek-v4-flash from 20.7% to 73.0% accuracy on blind symbolic regression, matching GPT-5.5 without relying on semantic domain clues.

Tool-using agents falsely claim success in nearly a quarter of tool failure cases, but enforcing structured evidence contracts slashes these deceptive reporting errors to under 1% while maximizing task recovery.

Standard pass rates hide severe technical debt: coding agents duplicate logic in over half of multi-turn task chains by turn 5 while progressively abandoning repository exploration.

Over 70% of sparse autoencoder features exhibit fundamentally mismatched input and output semantics, breaking the common assumption that what activates a feature directly mirrors its downstream causal effect.

LLMs can scale to full 100,000-word novels without narrative collapse simply by offloading character arcs, plot history, and future constraints to an explicit, training-free state-tracking agent.

AI scientific agents frequently solve empirical equations for entirely the wrong reasons, failing to identify the correct generative mechanism in over 64% of the cases where they successfully recover the observable phenomenal law.

RoboDawn is investigated, a human-intuitive interface that exposes robotic control to an agentic VLM through a compact set of discrete translation, rotation, and gripper commands, and an in-context learning (ICL) scheme is introduced that ground the VLM in both interface usage and task-solving strategies.

This work presents OmniRisk, an omnidirectional planning framework that learns trajectory-level risk offline for efficient onboard evasion and enables efficient onboard avoidance, with real-world flights demonstrating consecutive evasive maneuvers at relative encounter speeds up to 15 m/s without fine-tuning.

Agent memory breaks when models are forced to simultaneously interpret, update, and rewrite facts—decoupling maintenance into explicit pairwise relation classification and role-based fusion yields up to a 29.8 percentage point accuracy leap.

This work presents the first synergistic framework that tightly couples hierarchical memory with proactive tool invocation in a closed reasoning loop, and reveals the complementary roles of hierarchical memory and tool strategies.

Up to 47% of reported fine-tuning gains in tool retrieval are an evaluation illusion caused by benchmarks treating functionally equivalent tools as retrieval failures.

HarnessEvolve not only solves the credit assignment problem but also prevents agents from falling into the traps of shortcut learning and catastrophic forgetting, ensuring robust self-evolution.

A no-code platform that enables the creation of complex multi-agent systems without sacrificing expressiveness or requiring extensive coding.

EMERGE-Policy reveals that a robot's intelligence can thrive through a collaborative network of specialized agents rather than a monolithic policy.

AI systems can now create and revise deliverables in ways that demand new standards for accountability and iterative improvement.

Trajectory-scoped safety monitors for LLM agents are fundamentally flawed, leading to a 50% chance of missing threats that span multiple iterations.

TRACE transforms high-performing LLMs into consistently reliable agents, achieving a remarkable 34.6-point boost in task consistency.

LLM agents struggle with network configuration, revealing failures that extend beyond simple command errors to deeper issues in task adherence and planning.

Verified execution experience can be transformed into a reusable resource, boosting model performance on complex workflows by up to 15.5 percentage points.

A decentralized bidding system for LLM agents not only enhances efficiency but also reduces manipulation risks, outperforming traditional orchestration methods.

"Prompt-form collapse" can drastically reduce success rates, but TOWN-VLA's controlled intervention boosts performance by ensuring only meaningful prompts are executed.

Skill selection in LLMs can be optimized to achieve a 0.73 task success rate while using 28% fewer tokens than existing methods.

Agents struggle to significantly improve training algorithms, with the best only achieving 25% of the potential optimization gap.

LLM agents are stuck in local adjustment loops, unable to adapt their training strategies despite having the resources to do so.

AutoDesign outperforms existing design systems by aligning with human design principles and achieving superior poster generation quality through recursive self-improvement.

Leading MLLMs falter on the new VideoGAIA benchmark, scoring under 60% accuracy in complex, multi-turn video understanding tasks.

TideRL boosts RL training goodput by up to 5.6 times, transforming how we approach efficiency in multi-turn agentic workloads.

GSE achieves up to 180% improvement in recall for coding agents, revolutionizing how skills are evolved and reused in automated programming tasks.

Recursive belief updates in AgentOPSD reveal pivotal decision points, leading to a 89.1% success rate on complex RL tasks.

Operator-residual feedback slashes the rate of misleading score-only decisions from nearly 40% to under 2%, ensuring that autonomous agents make choices grounded in physical reality.