Alibaba DAMO Academy

MLLMs can now natively weave elastic visual representations, adaptively allocating detail to save 43% of tokens while preserving near-perfect performance.

Clinical LLM agents can now autonomously gather evidence from longitudinal medical records and imaging, achieving superior accuracy to GPT-5.4 and Claude Opus.

Generative world action models can run up to 3× faster in closed-loop control without fine-tuning or task degradation by dynamically repairing and reusing intermediate computational states instead of recomputing them from scratch.

Treating all latent tokens identically bottlenecks multimodal test-time compute—disentangling hidden-state updates into visual-grounding and high-entropy reasoning channels beats standard output-space search under matched sample budgets.

Uniform token-level distillation often punishes valid student reasoning, but dynamically reallocating teacher supervision using verified outcome agreement unlocks immediate gains across math and code tasks without relying on hard trajectory filtering.

Personal GUI agents waste most of their retrieved history; this framework activates experience only when it's actually applicable, boosting performance and efficiency.

Agent memory can be compressed by 50% with virtually zero performance degradation (retaining up to 99.7% accuracy) and a 2x retrieval speedup by structuring historical context into event-centric maximum spanning trees.

ARISE-RL transforms agent training by enabling robust self-evolution through a novel rubric-mediated co-evolution framework, achieving state-of-the-art performance across diverse tasks.

Agents can be misled into premature conclusions, but a new method reveals how to alleviate this pressure and improve decision-making.

SkillZip Pro slashes skill bundle size by 38% without sacrificing performance, revolutionizing how self-evolving agents manage their execution paths.

Achieving 94.52% of state-of-the-art performance with just 25% of the training data reveals a groundbreaking efficiency in visual instruction tuning.

Achieving synchronized audio-video generation at 2K resolution with a compact 7B model could revolutionize content creation and accessibility in multimedia applications.

Tracking both task advantage and actual reward gains can drastically improve the efficiency of multi-task reinforcement learning for LLMs.

NavMCP transforms isolated navigation rollouts into a cohesive, persistent interaction framework, leading to a 14.9 percentage point improvement over traditional episodic interfaces on long-horizon tasks.

LoopArena reveals that even the best models struggle with long-horizon task guidance, achieving only a 24.69% success rate in complex coding scenarios.

LiveSim transforms user behavior simulation by dynamically adapting to the evolving interactions in live-streaming environments, leading to unprecedented accuracy in risk analysis.

Relative calibration can significantly enhance the evaluation of memory consistency in video world models, revealing that traditional metrics may overlook critical distinctions.

HMGCLIP achieves superior attribute discrimination in e-commerce by leveraging a novel hypergraph structure that captures both fine and coarse-grained product semantics.

Continuous skill verification in RL agents leads to a substantial performance boost, outperforming traditional static skill banks.

Trigger items can dramatically enhance recommendation relevance, and CRRN leverages this to outperform existing methods in CTR prediction.

RecVerse outperforms traditional simulators by maintaining a nuanced understanding of user intent and memory, leading to more realistic shopping behaviors.

A single agentic framework can seamlessly tackle multiple intelligent document processing tasks, outperforming specialized models in the process.

Bridging fragmented research, this survey offers a unified taxonomy that redefines human-centric intelligence in the age of foundation models.

Generative supervision can transform the landscape of moiré removal, yielding substantial performance gains even in the most challenging real-world scenarios.

EditBridge achieves ultra-high-resolution image editing at 4K with a speedup of up to 8.4 times, while maintaining high fidelity and perceptual quality.

LENS achieves 84.8% evidence recall without the overhead of pre-materializing evidence, redefining how we approach dynamic document retrieval.

CPI-Bench reveals significant performance gaps among image editing models, offering a more nuanced evaluation that aligns with real-world user experiences.

Realistic predictions in robotic manipulation now come with precise arm control, ensuring the right actions yield the expected outcomes.

HounsWorld achieves state-of-the-art performance in multimodal patient-state inference by seamlessly integrating CT imaging and clinical language, transforming how we interpret medical data.

Nearly 20% of LLM agent violations occur even after agents acknowledge safety constraints, highlighting a critical gap in execution awareness.

Released tokenizer vocabularies can yield precise estimates of hidden corpus compositions, revealing insights that were previously obscured.

Emotion-driven feedback can transform multi-turn dialogue systems, boosting empathetic responses and model performance significantly.

REST achieves few-step image generation that rivals traditional 40-step methods while slashing training costs by over 75%.

Task learnability can significantly enhance RL efficiency in LLMs, leading to better performance with less data.

BCSD enables LLM agents to leverage external skills more effectively, achieving superior performance by integrating dual-context evaluations.

BOUND achieves up to 5.6 EM points improvement over traditional methods by correcting local search-control errors in real-time.

C2C-Explorer boosts LLM inference efficiency, achieving a 44.1% increase in goodput and a staggering 98.4% reduction in memory usage.

Event presence can mislead interpretations of skill execution fidelity, revealing critical discrepancies in agent performance assessments.

FullDiT not only outperforms leading commercial music generators but also redefines how we approach music rendering by leveraging full-context generation techniques.

CEDAR transforms demand forecasting by enabling accurate simulations of future sales outcomes based on planned actions, outperforming traditional methods that fail to account for decision-driven changes.

Real-time character animation is now feasible with Wan-Animate-2, which achieves high-fidelity results without the pitfalls of traditional motion representation methods.

Reducing token overhead by pruning redundant communication edges allows multi-agent systems to achieve better performance without the computational burden.

Output extrapolation in STEP-OPD enables a unified student model to surpass the performance of its specialized teachers across all evaluated tasks.

Current instruction-based video editing models are far from satisfactory, revealing critical gaps in evaluation that could reshape the field.

PrivDPO achieves robust LLM alignment while maintaining privacy, outperforming traditional methods in balancing privacy and utility.

SFC redefines semantic understanding in spoken language tasks, achieving superior accuracy and adaptability in open-domain contexts.

Sparse rewards can be effectively optimized without losing the advantages of fine-grained feedback through a novel two-stage training approach.

OmniPack achieves a remarkable 98% performance retention with a staggering 83.3% reduction in computational load, revolutionizing token compression for omni-modal models.

Token-level credit assignment in multi-turn language agents can be effectively enhanced by integrating teacher preferences through a novel self-distillation approach.

A new benchmark reveals that AI models can be rigorously evaluated on deep research tasks, highlighting their nuanced capabilities across diverse topics.

Visual identity discrimination is the missing link in Universal Multimodal Embeddings, and this work provides a robust framework to integrate it seamlessly.

VC-Tooler achieves unprecedented adaptability in visual tool use, outperforming existing models by 60% in agentic reasoning tasks.

AdaThinkV achieves 40.79% accuracy in video reasoning while using 22.7% fewer tokens than its strongest adaptive baseline, showcasing a breakthrough in token-efficient reasoning.

Follow-up suggestions for image editing can be made 32.70% more effective by integrating multimodal feedback into conversational systems.

Solution Hacking reveals that up to 44.1% of answers from frontier LLMs may be misleadingly credited as correct due to invalid reasoning shortcuts.

Misalignment between visual evidence and predicted timestamps in video grounding can lead to substantial performance drops, but CAVE effectively bridges this gap with boundary-specific rewards.

Synthesized egocentric videos can elevate real-robot task success rates by over 10% through enhanced training data.

LoopMemGR transforms generative recommendation by integrating past recommendations into the memory framework, leading to more informed and effective predictions.

ReAlloc not only optimizes marketing budget allocation but also achieves simultaneous lifts in pay order and income, challenging traditional approaches in e-commerce.

SERPO achieves up to 20.63 points improvement on HealthBench by enabling language models to self-evolve their evaluation criteria in real-time without external feedback.

LLMs can identify flagged security issues but struggle to uncover silent intrusions and create effective remediation plans, revealing a critical gap in their utility for real-world incident response.

Context-aware routing and compression in SPARC leads to superior generative recommendations by preserving crucial information without inflating input size.

TopoGR reveals that preserving the topology of semantic IDs can dramatically enhance item relatedness in generative recommendation systems, leading to superior performance.

SWAG-Bid redefines auto-bidding by integrating long-term market forecasting and adaptive control, leading to superior advertising outcomes.

Evolving solver and rubric skills in tandem reveals hidden weaknesses and boosts performance by up to 5% without relying on fixed evaluation criteria.

MTGuard effectively reduces harmful tool use in LLM agents, combining static and dynamic analysis to enhance security without sacrificing performance.

OmniDelta achieves a 1.64x speedup in inference while reducing GPU memory usage by 22% at just 25% token retention, redefining efficiency in OmniLLMs.

Achieving real-time audio-video generation without sacrificing visual fidelity or synchronization, TaoMate redefines the limits of digital human technology.

Achieving a 60x speedup in avatar generation without compromising visual quality could revolutionize production workflows in multimedia content creation.

TreeAdapter revolutionizes species image generation by harnessing hierarchical taxonomic structures, achieving unprecedented detail and accuracy in fine-grained outputs.

Achieving superior auto-bidding performance with less than 10% of the fine-tuning parameters required by traditional methods could revolutionize how LLMs are applied in dynamic auction systems.

Repeated revisions in coding agents can lead to a significant drop in reliability, highlighting the need for structured evidence-bound contracts in code repair processes.

OmniScope reveals that treating audio and video relevance separately can drastically enhance performance in omnimodal models, achieving remarkable efficiency gains without sacrificing accuracy.

Current SID-based generative recommendation models can sometimes reach future items but falter with completely unseen tokens, revealing critical gaps in cold-start handling.

VLM-IE3D achieves state-of-the-art performance in 3D tasks by seamlessly integrating implicit and explicit geometric representations from RGB inputs.

ECoM Reasoning boosts spoken language model accuracy by 21% while slashing token usage to just 40% of the standard approach.

TSGR redefines e-commerce search by ensuring that high-value items are prioritized in retrieval, achieving measurable business impact in a competitive market.

Open-Vocabulary Gaze Object Prediction can now effectively recognize and localize objects from a diverse array of unseen categories, outperforming traditional methods.

Achieving high visual fidelity while ensuring physical consistency in video generation could redefine the standards for simulating realistic interactions in AI-generated content.

Achieving up to 17× faster inference for long-context LLMs without compromising output quality could redefine efficiency standards in AI applications.

CODA redefines edge video diffusion by achieving up to 1.80x faster inference and 1.74x better energy efficiency without sacrificing output quality.

TmallGS redefines e-commerce search ranking by optimizing feature representation and interaction, resulting in substantial performance gains over traditional models.

Aesthetic assessments can be dramatically improved by focusing on key moments and endings, as shown by Peak-End-Net's state-of-the-art results in video evaluation.

Incorporating discount rates into conversion predictions can boost online sales performance by over 3% in real-world applications.

Over 60% reduction in post-purchase redundancy reveals a critical flaw in how traditional recommendation systems interpret user intent.

A structured evidence-state approach can boost omni-modal QA accuracy by over 30%, transforming how agents gather and validate information across diverse sources.

Bias in LLMs can be traced to specific geometric structures in their hidden states, allowing for precise control over scoring outcomes.

AWA-RL boosts precision in search agents by up to 10.3% by penalizing hallucinations, challenging the status quo of LLM training.