Steady Thought (ST) introduces a thought-level preference optimization to fix under-thinking by enforcing single-path continuations at switching points and preferring new responses, reducing token use while maintaining accuracy across models and domains.
The document outlines a ASD-STE100-inspired prompt engineering framework that escalates output formats from text to interactive HTML or video, targeting Claude and other advanced models. It prescribes concise, controlled language rules, terminology consistency, and structured, stepwise explanations to enable high-clarity, multi-format artifacts.
The piece argues that many daily prompting practices are technique-based, not named, and offers a practical, example-filled overview of ten prompting methods (Zero-Shot, Few-Shot, Chain-of-Thought, Tree-of-Thought, etc.) used in deep learning contexts.
TAC-CoT introduces a tool-empowered implicit chain-of-thought that interleaves latent reasoning with explicit tool calls, boosting accuracy and reducing computation while maintaining a usable interface for tool use.
KACE introduces Knowledge-Adaptive Context Engineering to optimize problem-solving prompts by selective evidence and tiered self-consistency; it outperforms several baselines on multiple math benchmarks and uses a store of domain-specific cards to guide reading.
Gemini thinking models reason internally before responding to complex tasks. The documentation explains how the Interactions API reports reasoning through thought steps and gives request examples in several programming languages.
The document describes two ideas for speculative decoding: parallel on-policy draft training and leveraging pretrained MTP to improve initial drafting. It outlines a process where, if the first draft is rejected, later drafts in the same round cannot proceed, and explains how MTP starts while DFlash continues, with DFlash drafting multiple tokens in parallel starting from T1 proposed by pretrained MTP.
LongCapHarness improves long-context reasoning through recursive updates to model capabilities and control logic, without data synthesis or post-training. Tests on Llama3-8B and Qwen2.5-3B outperformed recent data-synthesis-based baselines on HELMET and RULER at lower estimated cost.
The author shares a stance on using AI for structured tasks, preferring to handle planning while avoiding execution, and plans to spot-check AI outputs for a day before full trust.
This study tests reflective text-space optimizers GEPA and SkillOpt across two model setups on HotpotQA and MATH-500. Results show gains are task- and method-dependent, with improvements on MATH-500 but not on HotpotQA, and validation gains can reverse on held-out data.