w$^{2}$VLA introduces a modular vision-language-action model that decouples declarative and procedural knowledge. By restructuring information flow, it enables robust behavior cloning and zero-shot skill transfer to novel, dissimilar objects.
Mem-GF introduces a memory-efficient graph filtering method that approximates polynomial graph filters using Krylov subspaces, avoiding storage of the full item similarity graph. It achieves up to 5.74× lower memory usage and 4.38× faster runtime while outperforming state-of-the-art methods in accuracy and scaling to datasets with tens of millions of interactions.
A unified zero-shot framework, ZeProM, uses a pre-trained Video-Language Model to jointly perform procedural mistake detection and temporal action segmentation. It achieves up to 4.4 point improvement in EDA and 2.0 point in F1@.5 on EgoPER tasks, matching or exceeding supervised methods without task-specific training.
A benchmark of 8 LLMs on 300 synthetic doctor-patient dialogues found 12 high-impact hallucinations and 520 clinically relevant omissions. Omissions were far more common than hallucinations, with DeepSeek excelling in prose and cost but missing many safety facts, while Claude Opus had fewest omissions but poorer prose quality.
VibeThinker, a 3-billion-parameter language model, outperforms Opus 4.5 in reasoning tasks using a novel SFT+GRPO training approach. The model was introduced in a paper available on arXiv, with details shared in a Reddit post.
Baidu has introduced a new parsing model called One-shot Long-horizon Parsing. The model enables efficient, long-range understanding of text with minimal training data, as demonstrated in a GitHub repository.
GPT-5 Pro provided key insights into T cell behavior, resolving a 3-year-old immunology puzzle. The discovery may advance research in cancer and autoimmune diseases.
The creator of EverFern asks which local models work best for agentic workflows and browser/computer use. They note that model intelligence is rarely the bottleneck, with reliability and recovery systems being more critical than model choice.
A human evaluation on Design Arena's leaderboard reveals GLM-5.2 performs nearly as well as Fable 5 in game development tasks, placing just one step below it. The model, based on open weights and MIT licensing, is assessed as equivalent in capability to the best available Claude models, suggesting that standardized benchmarks may no longer accurately reflect real-world performance.
A user asks whether supervised fine-tuning (SFT) followed by reinforcement learning (RL) is still recommended for training Qwen 3.5 4B or 9B agents for multi-tool use, or if RL-only approaches yield better results. The post also seeks guidance on reward design and handling parallel tool execution in agent workflows.
Group-Graph Policy Optimization (G2PO) introduces a graph-based approach to enhance long-horizon agentic reinforcement learning by transforming interaction trajectories into state-transition graphs. It enables group-aggregated state-value estimation and edge-centric advantage calculation, improving credit assignment and reducing variance, and achieves up to 22.2% success rate improvement over GRPO on WebShop, ALFWorld, and AppWorld benchmarks.
Unlimited OCR introduces Reference Sliding Window Attention (R-SWA) to emulate human working memory, enabling long-document transcription without growing memory usage. By replacing decoder attention layers in DeepSeek OCR, it maintains a constant KV cache and achieves full document processing in a single forward pass under 32K token limits. R-SWA is also applicable to ASR and translation tasks.
A new Dual-Track Framework decouples template formatting from document processing by using an offline track to extract template constraints into a reusable manifest and an online track with a hybrid pipeline. It limits LLM use to reasoning tasks like metadata and bibliographic handling, while applying rule-based engines for deterministic operations, improving structural fidelity, layout compliance, and compilation success over baseline methods.
ToolGraph enhances multi-turn tool-using agents by integrating schema topology, transition weights, and history-aware controls. Training with DPO on 161 divergence-point preference pairs improves performance: ToolGraph+DPO achieves a 16.8% relative reward gain over baseline, especially in airline and retail tasks, with reward positivity emerging as the key diagnostic signal.
PRIDE introduces a knowledge distillation method that transfers empathetic reasoning from large models to smaller ones using privileged information available only during training. It achieves competitive or superior performance on empathy-related tasks by leveraging structured prompts, multi-source attention, and dual-alignment loss.
The discussion explores theoretical AI topics that have mathematical foundations and plausible implementation in current AI systems, such as large language models. Topics include reasoning chains, knowledge graphs, and probabilistic reasoning, all of which are grounded in formal math and show potential for real-world AI applications.
A study of 11 large language models across 21 disputed inventions shows that query language systematically influences which inventor is credited. Lower-status claimants appear more frequently when questions are phrased in their native language, while dominant Anglophone figures remain consistent. The findings suggest language acts as a switch that activates distinct national versions of history, indicating that LLMs function as systems of cultural memory.
DART enables hybrid reasoning models to route queries between direct answering and extended thinking without training data. It uses two no-think drafts to decide response mode and estimates thinking budget from draft disagreement. DART improves accuracy by up to 9.0 points in math and 22.-5 points in code reasoning while reducing thinking tokens by 15-69% and 51-63% respectively.
Researchers identify Memory Contagion, a phenomenon where evaluator bias propagates across time in agent memory. Even with perfect memory consolidation, bias spreads to future agents retrieving from the same memory store, with contamination detected as low as p=0.2. The effect varies by bias type: length bias is attenuated, while authority bias is amplified, indicating a bias-dependent interaction.
A study examines when intrinsic self-correction works by analyzing its performance across different task structures. The research finds that self-correction yields consistent gains only when the task supports explicit constraint verification, complex reasoning revision, or strategy evaluation. The results show SC is effective only in specific task contexts, not universally.