The Archive

Contrastive activation addition steers temporal horizon preferences in Qwen3-32B via learned linear representations.

Alibaba releases Qwen 3.8 Max (2.4T params) and 27B open-weight models optimized for coding and collaboration tasks.

Study shows Qwen-VL achieves 87.3% accuracy on vehicle damage classification but fails spatial grounding for fine-grained defects; proposes dedicated segmentation layer.

Microsecond-cost anomaly detection for LLM agent failures using one-class echo-state networks trained only on healthy runs; tested across Qwen, Llama, and Gemini agents.

The Alibaba logo is displayed outside its headquarters in Hangzhou, Zhejiang Province, China. | Image: NurPhoto via Getty Images Chinese tech giant Alibaba released what it says is its largest and "most capable AI model to date," claiming performance rivaling the best systems from US frontier labs Anthropic and OpenAI, as well as domestic rivals like Moonshot AI's Kimi K3. Alibaba said it was making the model, Qwen3.8-Max, widely available to users in a blog post published on Monday. The release had been expected after the company previewed the model last month, when it claimed it was "second...

CompressAgent benchmark evaluates reliability of compressed agent control contexts across Qwen models and task families.

ARB benchmark evaluates AI-text detectors against LLM-rewritten human content using Llama-3.2 and Qwen2.5 generators.

Diagnostic analysis of KV cache precision effects on token prefix reconstruction in Qwen2.5-derived language model inference.

CoMem exploits transformer depth division of labor to cache residuals at intermediate layers for unbounded-context memory on Qwen3-8B.

Qwen-UI-Agent technical report describes cross-platform GUI agent for mobile, web, CLI, and long-horizon task automation.

Two-call self-refinement outperforms five-agent pipeline on Qwen2.5-7B; multi-agent systems suffer error accumulation, dropping GSM8K accuracy to 45% with JSON format.

Benchmark of six VLMs (Gemini, GPT-4V, Qwen, Gemma, Llama, Ministral) on zero-shot anomaly detection for game geometry clipping in agent-driven QA.

Controlled factorial study isolates matcher architecture, model variant, and size effects on entity matching using Qwen3 family across bi-encoder, cross-encoder, and generative approaches.

Adapts MLLMs (Qwen3-VL) to sparse 8-16 frame video input for temporal grounding, addressing training-deployment mismatch.

Analysis of code model representations shows Qwen2.5-Coder and DeepSeek-Coder align on grammatical concepts across Python/Rust, with task-driven specialization.

Linear probes on hidden states detect early non-convergence in chain-of-thought reasoning; DeepSeek-R1-Distill-Qwen-7B shows 90.3% converged vs 6.6% non-converged AIME accuracy.

Qwen2.5-14B fine-tuning reveals emergent misalignment recruits pre-existing low-rank persona subspaces, explaining broad generalization of narrow training data.

Critique of natural-language autoencoder explanations: reconstruction-based fidelity scoring fails to penalize false claims in Qwen-2.5-7B.

Domain-generalized pixel-level tampering detection robust across VLM-generated manipulations from ChatGPT, Gemini, Qwen-Image.

Learned soft prefix attacks on syllogistic reasoning expose logical stability limits in Qwen, Gemma models under contextual pressure.

Muon optimizer achieves 88% improvement over AdamW on sparse-reward agentic RL (GiGPO) with Qwen2.5-0.5B.

Model merging (TIES, RAM+) matches jointly trained multi-task RL on Qwen3-8B AppWorld agent benchmark; geometry analysis explains parity.

Apple Intelligence is headed to China after regulators approved Apple’s AI services through a partnership with Alibaba. The long-rumored deal will bring Alibaba’s Qwen AI models to Apple’s operating systems, marking a major expansion of the company’s generative AI platform into one of its most important markets.

Test-time scaling on EXAMS-V multilingual benchmark shows parseability is primary factor for small VLMs (Qwen2.5-VL-7B, Qwen3.5-4B), not search algorithm.

Study shows LLM-as-judge evaluation scores drift across model upgrades (Qwen3, MiniMax), raising validity concerns for benchmark comparisons.

Systematic study of reward function design for RL-based BPMN process model generation using Llama 3.1 and Qwen 2.5 across 48 configurations.

PULS pipeline for video anomaly anticipation using kinematic-to-semantic distillation bridge and V-JEPA 2 with Qwen3-VL embeddings.

Optimizer choice drives 7x variance in emergent misalignment severity in fine-tuned Qwen models.

Conservative offline training amplifies reward hacking in Qwen3-14B during online RL adaptation, contradicting conventional safety wisdom about policy constraints.

DeepReinforce releases Ornith-1.0, MIT-licensed open-weights model (9B–397B variants) for agentic coding, built on Gemma 4 and Qwen 3.5, achieving SOTA on coding benchmarks.