Researchers introduce Skill Self-Play (Skill-SP), a reinforcement learning framework that addresses the diversity-vs-verifiability dilemma in LLM self-evolution by using agent skills as a middle ground. The system comprises a proposer, solver, and dynamic skill controller that co-evolve in a continuous loop: the proposer generates tasks conditioned on sampled skills, the solver explores solutions, and the skill controller updates an expanding skill library based on execution feedback. Evaluations on tool-use and reasoning benchmarks show consistent performance gains on capable backbones and recovery for initially misaligned models. Code is released under the Qwen-Applications GitHub organization, suggesting Alibaba/Qwen team involvement.
This arXiv paper investigates PortLLM, a training-free and data-free scheme for adapting LLMs after continual pretraining, extending prior short-term results to 10 continual pretraining steps across Mistral, Gemma, and Qwen base models. The authors find that LoRA patches remain portable across longer update horizons, suggesting repeated fine-tuning is unnecessary when base models are periodically updated. Two theoretical analyses are offered, identifying near-orthogonality of high-dimensional vectors as the geometric mechanism underlying temporal portability. The work has practical implications for reducing fine-tuning overhead in production deployments with frequently updated base models.
Researchers introduce DAIS (Dependency-Aware Intermediate QA Supervision), a training-time framework that converts teacher rationales into stage-level QA records where each intermediate step is conditioned on prior reasoning states. Evaluated on GDPR, AIACT, MedQA, and FOLIO benchmarks using Qwen backbones, DAIS outperforms answer-only, flat chain-of-thought, and independent-QA baselines, with up to 5.6% and average 4.2% gains on policy-compliance tasks. Ablations confirm that dependency conditioning contributes beyond simply adding more intermediate text, suggesting it as a lightweight auxiliary supervision signal.
Researchers present an MLIR-based compiler pipeline for deploying large language models on AI accelerators, using two dialect layers (TopOp for framework-agnostic graph representation and TpuOp for hardware-specific lowering). The method splits each Transformer layer into three static compilation stages (prefill, prefill_kv, decode) to handle the distinct computational profiles of prompt processing and autoregressive generation. The approach is implemented in the open-source TPU-MLIR compiler and LLM-TPU project, supporting Qwen, Llama, InternVL, and MiniCPM-V families with GPTQ, AWQ, and AutoRound quantization.
Researchers introduce ToolSciVer, a framework that equips vision-language models with three type-aware visual tools (table focus, chart-to-structure parsing, high-resolution zoom) to verify scientific claims grounded in figures, tables, and charts from papers. The policy is trained using Group Relative Policy Optimization (GRPO) with a composite reward covering correctness, format, tool-use efficiency, and validity. Experiments across five VLMs from three model families (Qwen, InternVL, Gemma) on SciVer and MuSciClaims benchmarks show improvements over prompting-based and RL-based baselines. The work is notable as the first tool-augmented framework specifically targeting multimodal scientific claim verification.
A newsletter digest covers four notable AI developments: PrismML (a Caltech/Khosla spinout) compressed Alibaba's Qwen 27B model to under 4 GB via ternary/binary quantization for on-device iPhone inference; Cognition released SWE-1.7 (trained on Kimi K2.7), jumping from 9.4% to 42.3% on FrontierCode 1.1 Main with novel RL and infrastructure techniques; Nvidia introduced Audex, a 30B unified audio-text transformer trained on 157B audio tokens; and Anthropic published research showing Claude's expressed values shift measurably by language across 309,815 conversations. Each item represents a distinct technical development across on-device inference, coding agents, multimodal models, and model behavior analysis.
Researchers introduce the One-Word Census, a minimal 31-prompt instrument that probes which one-word answers language models select from open-ended categories, applied to 44 models. Convergence is extreme — 41% of models chose 'serendipity' when asked to pick any word — yet conformity varies fourfold across models in structured ways: persona- and community-tuned models diverge most, while newest mainline flagships conform most. Within four model lineages (Claude, GPT, Qwen, Grok), conformity rises with each generation but reverses for the latest Claude and GPT flagships, suggesting possible repositioning. The field is more lexically concentrated than human norms in 18 of 20 shared categories.
A new arXiv paper analyzes why post-hoc linearization of causal self-attention degrades model quality, identifying key-dependent rank-1 orthogonal projections as the mechanism softmax relies on and explaining why delta-style networks outperform gated accumulation. The authors introduce structural interventions—sink tokens, short convolutions, and fixed-budget cache routing—applied in a frozen-backbone regime. Scaling across LLaMA and Qwen models up to 32B parameters, the approach outperforms prior post-hoc linearization baselines on MMLU and matches long-context retrieval of adaptive-caching frameworks.
A multi-story digest covers five distinct AI developments: ByteDance and Alibaba are shutting down customizable humanlike AI agents ahead of China's July 15 Interim Measures for AI-Based Anthropomorphic Interactive Services; Google released DiffusionGemma, an experimental 26B MoE diffusion-based text model generating 256-token blocks at 1,000+ tokens/sec on H100; Anthropic published findings from 400,000 Claude Code sessions showing domain expertise—not coding skill—drives agentic output volume; Seedance released version 2.5 of its video generator with higher resolution and longer clips; and Arena.ai expanded Code Arena to fullstack web development evaluation. The China regulatory action is the most significant item, representing a concrete enforcement moment for AI persona/companion regulation.
A new arXiv preprint formalizes the residual-stream gauge symmetry of transformer architectures, showing that RMSNorm models have a signed-permutation gauge group B_d = S_d ⋉ {±1}^d rather than the permutation-only S_d of LayerNorm models. The authors introduce sign-marginalized Hungarian matching and demonstrate that coordinate-preserving transport along fine-tuning trajectories recovers 91.1% of cross-run coordinates versus 60.3% for endpoint matching. Practical consequences include dramatically improved sparse autoencoder reconstruction (NMSE 0.004 vs 1.08), preserved steering vector effects, and correct AdamW optimizer state transfer — with implications for mechanistic interpretability, model merging, and activation engineering.
A new arXiv paper from Facebook Research evaluates whether mechanistic interpretability findings from open-weight models transfer to closed API-only models across prediction, attribution, and representation levels. Studying eleven models across four families (Llama, Qwen, GPT, Gemini), the authors find that prediction-level agreement substantially overstates attribution fidelity — models that agree on answers often disagree on why. They document an 'access-validity inversion' where white-box signals like attention patterns are stable across models but weakly predictive of causal attributions, undermining the common practice of using open surrogates to explain closed systems.
A new arXiv preprint proposes a multi-agent system for automated disinformation detection that emulates human annotator decision-making through consensus mechanisms, cognitive diversity, and hierarchical structure. The system uses open-source models (LLaMA, Kimi, Qwen, DeepSeek, LLaMA-Nemotron) and is evaluated on English, Polish, Slovak, and Bulgarian datasets across three fact-checking tasks. Results claim superior performance over individual LLMs including GPT-4 and GPT-3.5, with transparency benefits from using open weights models.
Researchers introduce MESA, a label-free framework for prioritizing security-critical communication edges in multi-agent systems (MAS) before attacks are observed. The framework combines six graph-theoretic metrics with two dynamic probes (ablation and masking) to rank edges by compromise risk, without requiring attack traces. Evaluated across three MAS scenarios, eight network topologies, and five open-source LLMs, MESA achieves mean Spearman ρ=+0.60 correlation with empirical per-edge attack success, and monitoring the top 10% of ranked edges intercepts roughly 3x more successful attacks than random allocation. The work highlights that attack impact in MAS is highly concentrated — a single compromised edge can account for up to 75% of total attack success.
Qwen has released Qwen3-ASR-1.7B on Hugging Face, a 1.7B parameter automatic speech recognition model supporting multiple languages including Chinese (Mandarin and Cantonese), English, Arabic, German, French, and Spanish. The model uses the Qwen3 architecture and is released in safetensors format. This extends the Qwen3 model family into the speech domain.
Alibaba's Qwen team released Qwen3-ASR-0.6B on Hugging Face, a 0.6B parameter automatic speech recognition model supporting multiple languages including Chinese, English, Cantonese, Arabic, German, French, and Spanish. The model uses the Qwen3 architecture and is available in HuggingFace-native format with safetensors weights. This is a compact ASR model extending the Qwen3 family into speech recognition.
Qwen released Qwen3-ForcedAligner-0.6B on Hugging Face, a 0.6B parameter token-classification model tagged for ASR (automatic speech recognition) forced alignment tasks. The model supports multiple languages including Chinese, English, Cantonese, French, German, Italian, Japanese, and Korean. This is a specialized small model extending the Qwen3 family into speech-text alignment use cases.
Researchers introduce SafeVec, a white-box safety evaluation procedure that measures LLM safety from internal hidden-state representations rather than generated outputs. The method extracts layer-wise refusal directions from a safety-aligned reference model, identifies stable layers where safe and unsafe behaviors are separable, and scores target models via a calibrated 0-100 Refusal Alignment Score (RAS). Evaluated across Llama, Gemma, and Qwen model families, RAS distinguishes aligned from uncensored/abliterated variants and correlates with output-level attack success rates while being substantially faster than judge-based evaluation. The approach addresses key limitations of output-level safety evals: cost, judge sensitivity, and dependence on fixed question banks.
Researchers introduce the first public multi-modal dataset of 100 aligned audio-transcript pairs of Turkish scam and benign phone calls, evaluating seven LLMs (Gemini 2.5 Flash/Flash-Lite/Pro, GPT-4o, Qwen Max/Plus/Turbo) under three input conditions. Transcript-based inputs consistently outperform direct audio processing, while human-corrected and uncorrected transcripts perform comparably. The work addresses a gap in low-resource language safety research and highlights the need for linguistically inclusive fraud detection systems.
Qwen has released Qwen-AgentWorld-35B-A3B on Hugging Face, a 35B-parameter MoE model (3B active) built on the Qwen3.5 MoE architecture. The model is tagged for world-model and environment-simulation use cases, suggesting it is designed to simulate environments for agent training or evaluation. It is paired with a dataset called AgentWorldBench, indicating an associated evaluation suite. Early engagement is minimal (0 downloads, 4 likes) but the model represents a notable direction in agent-environment modeling from a major open-weights lab.
Researchers introduce KDoS (Knowledge Distribution-optimized Synthesis), a framework that uses a three-stage feedback mechanism guided by 'knowledge density' to optimize the distribution of synthetic training data for LLMs. Rather than stopping at preset token counts or fixed ratios, KDoS dynamically adjusts synthesis to avoid sparse or redundant domain coverage. Experiments across Qwen, Ling, and LLaMA models (0.6B–16B parameters) on 1B–5B token scales show consistent improvements over baselines on six knowledge benchmarks. A key finding is that an optimal knowledge distribution exists and remains stable across model families and scales.
RubricsTree is a new evaluation framework for LLM-powered personal health agents, built around a hierarchical taxonomy of over 100 clinically-verifiable Boolean rubrics derived from 4,000 real user queries and curated with physician oversight. A context-aware router activates only relevant rubrics per query, enabling scalable yet expert-aligned evaluation. The framework outperforms strong LLM-as-a-judge baselines on expert alignment and, when used as training signal, yields up to ~66% relative gains on HealthBench across Gemini, GPT, and Qwen model families. The work addresses a concrete bottleneck in clinical deployment of health AI: the cost-quality tradeoff in evaluation.
Researchers introduce RePro (Retrospective Progress-Aware Training), a framework addressing the gap between step-wise RL optimization and metacognitive task-progress awareness in LLM agents. The approach uses a forward-then-reflect rollout paradigm where agents execute actions online and then retrospectively assess step-wise progress given the completed trajectory and known outcome. Evaluated on WebShop, ALFWorld, and Sokoban, RePro achieves up to 12% absolute success rate gains over baseline Qwen-family models without requiring continuous external supervision.
Open Interpreter is an open-source Python coding agent framework supporting open-weight models including Deepseek, Kimi, and Qwen. The project has accumulated nearly 64,000 GitHub stars, with 45 new stars on the trending day. It provides a lightweight harness for running code-executing agents on locally-hosted or open models.
A new arXiv preprint proposes Reroute, a training-free plug-in that replaces the standard rank-and-remove visual token pruning paradigm in VLMs with a recoverable routing mechanism. Instead of permanently discarding low-ranked tokens, Reroute defers them to re-enter the candidate pool at later decoder stages, addressing the problem that token importance shifts across decoder depth. Evaluated on LLaVA-1.5 and Qwen backbones augmented with FastV, PDrop, and Nüwa pruning methods, Reroute improves grounding performance under aggressive token reduction without sacrificing general VQA accuracy. The approach preserves the theoretical compute and KV-cache budget of the underlying pruning method.
Qwen has released Qwen3.5-35B-A3B, a 35B-parameter mixture-of-experts image-text-to-text model with approximately 3B active parameters, published on Hugging Face. The model supports conversational use and is compatible with Azure deployment endpoints. With over 2.8 million downloads and 1,400+ likes, it has seen substantial community uptake.