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.

Current chemistry benchmarks conflate memorization with derivation and treat visual/textual inputs interchangeably; ChemArgus finally provides fine-grained diagnostic scores that pinpoint *where* a model's reasoning chain breaks.

Chain-of-thought token bloat can be slashed by 27% while actually boosting accuracy simply by recognizing that reasoning capacity lives almost entirely within the null space of a base model's dominant singular vectors.

Fysiverse-3D-Vision is proposed, a unified vision-language-geometry framework for generative 3D scene reconstruction and executable asset construction from a single image that establishes a shared representation where spatial reasoning and geometric reconstruction mutually enhance each other, allowing object layouts to be inferred beyond the constraints of individual asset generators.

A reward-compatible bounded mixing mechanism for $\gamma\mathrm{OPD}$ that balances verifiable outcome feedback with the discounted OPD advantage to move beyond purely teacher-dependent optimization is developed.

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.

Chain-of-thought context can be halved without sacrificing accuracy by using the geometric trajectory of hidden states to selectively compress exploratory reasoning steps into continuous latent tokens.

RuleMem's innovative approach to active memory management leads to a staggering 54.3% improvement in conversational accuracy by transforming dialogue history into reusable logical rules.

Placing reasoning traces before context in transformers can boost long-context reasoning accuracy by up to 100% while using exponentially less memory.

CoMPASS achieves a remarkable balance, enhancing molecular predictions by intelligently integrating LLM insights only when uncertainty is high, leading to improved accuracy without sacrificing reliability.

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

Reducing visual token exposure by leveraging selective retrieval can dramatically enhance the performance of 3D medical image question answering systems.

Agents using ParallelWorld can efficiently evaluate multiple future trajectories, leading to superior decision-making in complex environments.

Explicit strategy induction can dramatically enhance LLM performance, but its effectiveness varies widely across task types and configurations.

DIAG reshapes practice distribution to maximize informative supervision, leading to significantly improved reasoning performance in LLMs.

FormuEvo can accelerate solver performance by up to 5.5× by intelligently evolving MIP formulations through LLM-guided optimization.

Teacher rewards can mislead learning, but R2-OPD filters out conflicting feedback to boost reasoning performance in language models.

Group-Calibrated On-Policy Distillation boosts long-context reasoning performance by reconciling teacher guidance with verifier feedback, achieving up to a 12-point increase in benchmark scores.

SCOUT achieves a remarkable 16.85% improvement in spatial reasoning benchmarks, setting a new standard for Vision-Language Models.

FactorDrive redefines autonomous driving by seamlessly integrating spatial-physical evidence into adaptive reasoning, achieving state-of-the-art planning performance.

CED reveals that VLMs can be trained to prioritize evidence-based reasoning over language shortcuts, leading to more reliable visual understanding.

Grounded language comprehension, rather than free-form reasoning, is the key to unlocking superior performance in Vision-Language-Action models.

Argus achieves a 78% success rate on long-horizon reasoning tasks while using 21% fewer tokens in mature workflows, showcasing a revolutionary approach to agentic autonomy.

Existing unlearning methods can leak sensitive knowledge through multi-hop reasoning paths, exposing a critical vulnerability in LLMs.

Constraint-First Reasoning reveals that explicitly managing answer-space constraints can dramatically enhance the accuracy of mathematical problem-solving in language models.

LLMs struggle to effectively integrate and order evidence in attack chain reconstruction, with top models only succeeding 39.6% of the time on critical tasks.

The proposed Attention-Guided Switching method allows MLLMs to dynamically balance between visual fidelity and logical coherence, achieving unprecedented efficiency in reasoning tasks.

LedgerMind reveals that grounding multimodal reasoning in a structured evidence ledger can significantly mitigate common pitfalls like entity hallucination and unsupported reasoning.

GeoLens outperforms traditional single-tool approaches by effectively integrating multiple visual reasoning tools, achieving superior accuracy and efficiency in complex remote sensing tasks.

REFACT reduces token consumption while enhancing the density and faithfulness of reasoning traces in large language models, ensuring that every cited fact meaningfully supports the answer.

SLPO enables latent reasoning models to leverage outcome-reward learning, dramatically enhancing their performance and efficiency in complex reasoning tasks.

By intelligently deciding when to zoom in on audio-video content, OmniReasoner boosts reasoning accuracy while reducing computational costs, transforming how LLMs handle complex multimodal inputs.

MLLMs struggle to discern genuine consensus from pseudo-consensus in multi-party meetings, revealing critical gaps in their Theory of Mind reasoning abilities.

Continuous tracking of dynamic object evidence can transform MLLMs' ability to understand and interact with dynamic environments.

OpsMem's innovative dual-memory approach boosts failure diagnosis accuracy by leveraging operational experience, achieving unprecedented performance improvements over existing methods.

CheckRLM cuts error accumulation in reasoning chains by correcting factual inaccuracies in real-time, outperforming traditional approaches.

Visualizing code dependencies can dramatically enhance issue-resolution performance, outperforming traditional text-based navigation methods.

LLMs show significant vulnerability to logical fallacies, with distinct profiles of resilience that could inform future model training strategies.

CoLT slashes inference time by over 10x while enabling multi-modal models to reason more efficiently with fewer steps.

DRIFT not only sets a new state-of-the-art in self-improvement for language models but also redefines how we can dynamically adapt learning strategies based on problem difficulty.

Current T2I models fall short in scientific reasoning, but fine-tuning on the new SciIR dataset boosts performance by over 20%.

Evaluating LLM agents in microservice failure diagnosis reveals that traditional outcome-based benchmarks miss critical reasoning processes, which these new datasets effectively capture.

ProMSA achieves superior accuracy in KB-VQA by dynamically selecting retrieval strategies, outperforming traditional fixed pipelines.

SR-PPO achieves significant gains in reasoning tasks by effectively assigning credit to individual tokens from a single rollout, transforming how we approach reinforcement learning in language models.

PointVG-R achieves a groundbreaking 15.86-point boost in mIoU by integrating geometric reasoning into visual grounding tasks, reshaping our approach to spatial interpretation in models.