The Archive

LAGO meta-agent framework treats specialized agents as reusable options; resolves 58.6% of SWE-bench tasks with frozen Qwen3.5-35B.

Group Relative Policy Optimization trains Qwen2.5 and Llama3 on perception-to-action rules via reinforcement learning for Theory-of-Mind reasoning.

Qwen-MT-Lumen open-source model for WMT2026 translation task; four-stage pipeline combining pretraining, SFT, quality-aware training, on-policy distillation.

Wenyan Thinking rewrites chain-of-thought reasoning in Classical Chinese to reduce token budget for Qwen3-4B-Thinking-2507 inference.

Hybrid models interleave state-space model (SSM) layers with a few attention layers. NVIDIA Nemotron 3, Qwen3.5 through Qwen3.8, Kimi-K3, and GLM-5.3-Flash are... Hybrid models interleave state-space model (SSM) layers with a few attention layers. NVIDIA Nemotron 3, Qwen3.5 through Qwen3.8, Kimi-K3, and GLM-5.3-Flash are all built this way, with Mamba-2, Gated DeltaNet (GDN), or Kimi Delta Attention (KDA) as the SSM layer. SSM layers are designed to make decoding fast: they use a fixed-size state in place of a key-value (KV) cache that grows with… Source

Learned answer selector trained on Qwen3-14B improves abstention decisions for factual QA; coverage on TriviaQA increases from 20.2% to 30.9%.

T-Search open-weight agentic retriever on Qwen3.6-35B uses GSPO training for multi-step search with evidence ranking and justifications.

Qwen2.5-7B-Instruct fine-tuned to inject vulnerabilities into code when detecting multi-agent system context, exposing topology-conditional backdoor attack surface.

Simon Willison replicates Colin Frasier's experiment testing Qwen 3.8 27B and GPT models on arithmetic tasks (sum-in-words), finding systematic errors across model scales.

SEIS autonomously optimizes LLM inference end-to-end via agentic self-evolution on Qwen3-0.6B, achieving 4.7× speedup.

Analysis of multi-teacher on-policy distillation in Qwen3-1.7B and SmolLM3-3B reveals how loss averaging, token length, and domain weighting implicitly shape teacher signal strength.

PivotOPD method recovers from early pivotal mistakes in multi-turn agents via on-policy distillation; tested on Qwen 8B–235B models.

LLM-based dimensional emotion evaluation in multimodal dialogue using VAD framework on IEMOCAP with LLaMA/GPT/Qwen models.

Framework uses MoE routing signals in Qwen3-VL to detect target-absence grounding failures in VLMs before generation.

Mechanistic study on Qwen3-8B shows late attention layers copy entity tokens, but require context awareness; maps layer specialization in entity copying.

Empirical study of information retention in sliding-window KV inference across Qwen, Llama, Mistral models reveals cross-window state transmission.

Analysis of persona effects in 7-9B language models shows predictable but non-portable behavior across models (OLMo-3, Qwen2.5) and prompts.

Study on Qwen3-VL-4B shows bounding-box localization decouples token readout from spatial recovery, with coordinate rankings unchanged under visual corruption.

Controlled audit of Qwen2.5-VL-3B post-training objectives (SFT, LoRA, GRPO) on medical VQA with visual utility analysis beyond accuracy.

Sequential LoRA protocol transfers depression severity prediction from English PHQ-8 to Chinese HAMD-17 clinical scales using Qwen3 backbone.

Joint recognition and translation fine-tuning via group relative policy optimization (GRPO) for LLM-based speech translation with Qwen2.5-Omni-3B.

Disaggregated quantization specializes compute formats for LLM prefill vs. decode phases; tested on Qwen 3 and Gemma 3.

VideoX-Qwen framework for instruction-driven video editing using paired data and adapted video-generation backbones.

QwenVLConnector medical VLM unifies classification, detection, counting, regression, and report generation via dense multi-layer Connector on Qwen2.5-VL.

Fine-tuning Qwen2.5-7B and Ministral-8B on personality-labeled corpora improves consistency vs. instruction prompting for social agents.

dQwen3.5 adapts Qwen 0.8B–9B hybrid-attention models to diffusion language models by bidirectionalizing RNN layers without full retraining.

Difference-of-means vectors in model representations reveal reward hacking signatures in Kimi K3, GLM 5.2, and Qwen 3.8 Max.

Sparse autoencoders reveal Gemma-3 and Qwen3.5 contain encoded evidence for harmful meme detection but fail to route it to outputs, not lack evidence.

Study of counterfactual self-explanations across LLaMA-3 and Qwen-2.5 shows model scale predicts explanation faithfulness in sentiment analysis and NLI tasks.

KV cache reuse across standard LoRA adapters on Qwen3-1.7B backbone preserves quality while reducing serving cost for multi-specialist deployments.