Eikos is an open-source family of typed-decision models (4B and 27B parameters) released under MIT license for structured question-answering in global finance and trade. Each model returns calibrated probabilities for decision options in a single forward pass, with support for multiple GPU formats and Apple Silicon, plus complete training pipelines and evaluation tools.
Alex Zhang discusses how language model input/output shapes have remained static since ChatGPT, with harnesses designed around autoregressive models rather than vice versa. He argues that alternative model architectures with constrained output spaces—like Jev, which outputs values in [0,1]—could enable more efficient solutions for specific use cases and agent designs.
Fly Language Model (FLM) is a language model trained from scratch on a connectome-derived recurrent network inspired by fruit-fly brain wiring, allowing inspection of neural states during inference. The project combines anatomical wiring constraints with learned dynamics, tested on WikiText-2 and BabyLM datasets to isolate the computational value of the fly-derived architecture versus conventional baselines.
Expensive large language models like Claude Mythos outperform cheaper alternatives on complex, long-form tasks and fact retention, but smaller models like Claude Sonnet handle routine work well. Simple prompting techniques—being specific, providing facts, admitting uncertainty, showing examples, and breaking tasks into steps—can significantly narrow the performance gap.
Anthropic released Claude Opus 5.5, a new model performing at Fable 5.1 level while costing 40% less than Opus 5. The model features improved communication, agentic coding capabilities, and zero data retention options, with pricing at $4/$20 and faster speed options available.
A research project demonstrates running Qwen3-0.6B language model inference in Linux eBPF kernel space, executing 28 decoder layers with token generation in verified BPF programs while using fixed-point arithmetic and BPF maps for state management. The prototype achieves working forward passes at ~1.2 seconds per token, split between C code for tokenization and weight loading and BPF programs for matrix operations, attention, and token selection.
Synthetic Hospital is an open, synthetic longitudinal EHR benchmark built from public medical-education material with 1,268 patients and 5,602 encounters, designed to overcome privacy barriers while providing verifiable ground truth grounded in standard medical ontologies. Physician reviewers distinguished synthetic records from real charts at near-chance rates, and frontier language models achieve at best a severity-weighted F1 of 0.73 on patient problem list reconstruction, revealing significant gaps in clinical AI performance.
Language models can function as decision models by using next-token probabilities to select actions from candidate options without requiring decision-specific training. Qwen3.5-9B, tested against Jev 1.13.0, achieved competitive accuracy on multiple benchmarks including WebPRM and DeepSWE tasks, with the approach using rotated letter scores and log probability aggregation across candidate positions.
Researchers developed a method to extract hidden chain-of-thought reasoning from frontier language models like GPT-6 Astra by using a custom API tool, finding that externalized reasoning matches native performance and reveals systematic differences in how models structure intermediate reasoning across mathematics, science, and code tasks.
This paper establishes empirical scaling laws showing that language model loss follows power-law relationships with model size, dataset size, and compute, spanning over seven orders of magnitude. The research demonstrates that larger models are more sample-efficient and that optimal training involves large models on modest data, stopping before convergence.
MEF LLM Studio is an educational Windows application designed for beginners to understand how language models work through interactive explanations and hands-on experiments. Users can explore tokenization, training dynamics, model comparison, and text generation without writing code.
JEV-Star is a system that combines language-model planning with StarCraft II control, enabling both macro-level game strategy and micro-level unit management. The framework integrates the Astra planner with JEV action selection across multiple game configurations, achieving wins on hard micromanagement tasks and full-game macro control with realtime performance.
Researchers discovered that language model agents encode signals for memory management (compression and recall) in their hidden states before taking actions, indicating the models already represent when these operations are needed. They propose PaMER, a framework combining state-guided compression with evidence retrieval that reduces context consumption while maintaining task performance on long-horizon agent benchmarks.
LatentPort demonstrates cross-model transfer of recurrent inference state from a 4B to 9B Qwen language model without replaying the source context, using hybrid-state handoff combining translated attention KV cache with Gated DeltaNet persistent-state components. The approach achieves near-native performance with only a 0.076 nats/token excess loss on continuation tasks.
HySparse2 is a hybrid sparse attention architecture designed for long-context language models that improves efficiency through two-level KV sharing between self-decoder and cross-decoder components. It replaces block-level sparsity with token-level sparsity and enables prefill computation to exit early, reducing computational cost and KV-cache storage while maintaining performance on long-context retrieval and multi-turn agent tasks.
NVIDIA introduces Nemotron-H, a family of hybrid Mamba-Transformer language models (8B to 56B parameters) designed for efficient inference while maintaining competitive accuracy. The models achieve up to 3x faster inference than pure Transformers, with the 56B variant trained on 20 trillion tokens in FP8 precision and capable of supporting ~1-million-token context windows.
A tutorial demonstrates implementing Jev, a decision classification model, in 25 lines of Python using the Qwen3 language model to classify email inputs into categories like legitimate, spam, or phishing by extracting and normalizing token logits into probabilities.
Claude Opus 5.5 achieved the top ranking on the Artificial Analysis Intelligence Index and received a 20% price reduction. The model matches GPT-6 Astra performance on benchmarks like Terminal-Bench 4.0 and AutomationBench-AA while offering improved cache hit discounts.
A comparative study tested three context compaction strategies—FutureOS, OpenCode, and Codex—on their ability to retain information from agent sessions. FutureOS retained 83% of queryable information, significantly outperforming OpenCode (47%) and Codex (38%), with the key difference being that FutureOS preserves assistant prose while others compress it away. The analysis reveals that tool output dominates context volume but is rarely referenced, while the sparse assistant text is the primary source of follow-up questions.
Transformers are a neural network architecture introduced in 2017 that power modern AI models like GPT, Llama, and Gemini. They use self-attention mechanisms to predict the next token in sequences and process text through embedding, transformer blocks with attention layers, and output probability layers.