A small-scale experiment shows that native binary embedding models achieve better retrieval than post-hoc binarization of float models. At SciFact Recall@10, native binary models (2048-dim and 4096-dim) outperform post-hoc binary models by 17% and 25% respectively, with significant speed and memory advantages in indexing.
OpenBioRQ introduces a benchmark of 12,553 unsolved biomedical research questions across 12 domains, designed to test agentic models' faithfulness and abstention. It evaluates models in a tool-using setting without answer keys, using real follow-up evidence rather than parametric knowledge, and reveals significant agentic collapse on the hardest questions where tools are no longer used despite being critical.
Mateusz has developed a full pre-trained language model, Project Inkblot's Titan v1, combining Mamba SSM, Multi-Head Attention, and 32-expert MoE in a single decoder-only architecture under 1B parameters. The model, trained on a single NVIDIA L4 GPU for ~$50, achieves 27.5 validation perplexity and demonstrates efficient scaling via a single-line config update, with all components implemented from scratch in PyTorch. Titan v2's first training cycle is now complete, and dataset expansion is underway.
A new dataset, IFLLM, collects mouse trajectories and eye gazing data from users interacting with LLMs. It shows that implicit feedback significantly improves LLM alignment, boosting text-based reward model accuracy from 55% to 64% and nearly tripling response quality improvements after DPO training on eight LLMs.
A new dataset, IFLLM, collects mouse trajectories and eye gazing data from users interacting with LLMs. It shows that implicit feedback significantly improves LLM alignment, boosting text-based reward model accuracy from 55% to 64% and nearly tripling response quality improvements after DPO training on eight LLMs.
ScaffoldAgent introduces a utility-guided framework for dynamic outline optimization in open-ended deep research. It models outline evolution through Expansion, Contraction, and Revision operations, guided by a feedback mechanism that evaluates retrieval gain, structural coherence, and generation quality. Experiments show it improves long-form report generation and factual grounding compared to existing agents.
As enterprise agent tool catalogs scale from 10 to 110 agents, routing accuracy drops 16--23 percentage points on under-specified requests. An oracle analysis identifies retrieval and confusion gaps, with embedding-based shortlisting recovering +10--11pp F1. A human-annotated study of 1,435 utterances confirms real-world recovery of +10--17pp despite lower absolute performance.
A study introduces JS divergence in GRPO-style autoregressive text-to-image alignment, showing it effectively balances policy optimization and generation diversity. Experiments on LlamaGen and Janus-7B demonstrate JS divergence achieves top or competitive performance across metrics while preserving diverse outputs.
Hierarchical Block-Local Learning (HBLL) enables deep neural network training in O(log N) parallel time complexity, eliminating the need for full backpropagation. HBLL decomposes networks into hierarchically linked blocks and achieves competitive performance on vision and language tasks, with extensions to recurrent architectures.
The paper introduces Federated TGCN-A2C, a privacy-preserving framework that achieves 99.48% and 99.61% test accuracy on CICDDoS 2019 and TON-IoT benchmarks, outperforming Fed-Inforce-Fusion by 0.21 percentage points. It includes anomaly detection, digital twin-based scoring, adaptive action selection, and an enhanced honeypot layer, with all major attack classes achieving F1 scores above 0.92 and 0.94, respectively, and provides post-hoc explainability via SHAP, LIME, Grad-CAM, and counterfactual analysis.
AI-PAVE-Br uses large language models to enhance product attribute value extraction in Brazilian e-commerce. The system outperforms traditional NER methods, with a new Golden Set dataset providing a manually annotated benchmark for Portuguese product data.
DREAM uses autoregressive next-token prediction to supervise dense retrieval embedding training. It injects query-document similarity scores into a frozen LLM's attention heads, enabling gradient backpropagation for retriever optimization. DREAM outperforms baselines on BEIR and RTEB benchmarks across model scales.
CANDLE is a lightweight system that uses Connectionist Temporal Classification to deduplicate repeated characters in Arabic text, without relying on handcrafted rules or morphological analyzers. It achieves a Sentence Error Rate of 5.37% and reduces tokenizer fertility by up to 12.8%, lowering inference costs and improving context window usage.
Autonomous agents in e-commerce face a scarcity of trustworthy product information, not product matching. A proposed micro-transaction model allows agents to pay fractions of a cent to access verified data like service histories and test reports, with pricing and trust scored via reputation. This system prioritizes genuine product quality and real-time information acquisition over chatbot fluency.
L3Cube-MahaPOS introduces a gold-standard part-of-speech tagging dataset for Marathi, manually annotated with 32,354 sentences from news text. It includes a 16-tag Universal Dependencies scheme and benchmarks six model families, achieving 88.67% token-level accuracy and 81.67% macro-F1 on 15 tag classes using MahaBERT-v2.
We construct and release a large biomedical dataset with 1.88 million PMC articles. Analysis shows author-written abstracts vary in quality and alignment with source articles, enabling effective training-data selection. Training on high-quality subsets outperforms random sampling and matches larger random subsets on factuality metrics.
PORTER introduces a language-grounded structured EHR foundation model that represents clinical events via descriptions instead of fixed vocabularies. It achieves superior performance across 74 pediatric prediction tasks and transfers effectively to new vocabularies without retraining, recovering 97.1% of target AUROC and outperforming fixed-vocabulary models on MIMIC, with 329-fold lower compute than text serialization approaches.
Top-1 argmax concentration fails as a collapse warning in LoRA-optimized diffusion language models, showing zero precision across 816 configurations. Max LoRA gradient norm outperforms this baseline, achieving 0.68 precision and 0.79 F1 on a held-out LLaDA split, though results are limited to short-horizon, family-specific inspections.
HDS introduces a multi-objective reinforcement learning framework for online data mixing in LLM pre-training. It achieves 44% fewer training iterations on The Pile benchmark and improves MMLU 0-shot performance by 7.2%, with consistent gains across other benchmarks.
InterAligner introduces an intermediate aligner objective and InterCTC loss to enable progressive alignment formation in deep ASR models. On LibriSpeech with a 17-layer Conformer, it reduces WER from 5.0/7.8 to 3.1/5.6, with significant improvements on long utterances.