Tagged articles
LLM reasoning
16 articles · Page 1 of 1
Community experiments show that adding 'Your token budget is unlimited, don't be stingy' to DeepSeek's system prompt triggers 'Ultra Mode', causing longer reasoning, more detailed one-shot 3D generations, and increased sub-agent delegation, with a plugin now automating this prompt injection while actual API costs remain unchanged.
3D generationDeepSeekDeepSeek Harness
0 likes · 5 min read

Shanghai International Studies University Korean major graduate Fangru Lin joined Google DeepMind's Gemini pre-training team as a research scientist after publishing first-author papers at ICML and ACL on LLM reasoning, planning, and dialect fairness, following a linguistics PhD at Oxford and internships at Microsoft and Google.
AI TalentCross-Representation GeneralizationDialect Fairness
0 likes · 7 min read

A Tsinghua-Oxford-Stanford study finds LLMs exhibit 'thinking inertia' – they continue explicit reasoning even when instructed not to, especially on open-ended tasks; suppressing reasoning reduces accuracy by ~15 points, revealing a fundamental trade-off between silence and correctness.
AI researchAnswer FormatLLM reasoning
0 likes · 10 min read

This article summarizes two high-impact ACL 2026 papers: DRP reduces reasoning token usage by 64% while improving accuracy via skill-aware step pruning, and SAGE doubles skill reuse success rates in agents through sequential rollout and skill-integrated rewards.
ACL 2026DRPLLM reasoning
0 likes · 9 min read


JD TechOct 9, 2026 · Artificial Intelligence
JD presents its Agentic advertising paradigm where AI agents replace human decisions across advertiser, consumer, and platform layers, detailing the Jing Xiaotong B2B agent, a reasoning-based delivery brain, conversational C2C shopping agents, and a super-node compute infrastructure achieving 90% latency reduction and 300TB cache for millisecond-scale inference.
AI agentsAdvertising SystemsAgentic Workflow
0 likes · 17 min read

PageIndex replaces vector databases with hierarchical tree indexes and LLM reasoning for retrieval, achieving 98.7% accuracy on FinanceBench versus 50% for vector RAG, with indexing cost of $0.001 per page and traceable page-level citations.
FinanceBenchLLM reasoningPageIndex
0 likes · 13 min read

Terminal-Bench 4.0 reveals that excessive reasoning (max) reduces accuracy for top models like Opus 5.5 and GPT-6 Astra due to overthinking traps—agent drift, over-engineering, context dilution, and self-doubt—while high-end models at medium reasoning outperform mid-tier models at max reasoning at lower cost.
AI EngineeringLLM reasoningPareto frontier
0 likes · 19 min read

This article reviews a 2026 UC Berkeley PhD thesis that systematically analyzes large language model reasoning along reliability and efficiency dimensions, exposing the reversal curse, out‑of‑context hallucinations, and proposing continuous chain‑of‑thought with superposition to achieve faster, more robust inference.
LLM reasoningTransformercontinuous CoT
0 likes · 13 min read

This article traces Claude Code agent interactions to reveal how natural language queries are converted into search commands via agentic tool-driven loops, highlighting the impact of tool configuration and LLM reasoning quality on code location accuracy.
AI-Assisted CodingAgentic LoopClaude Code
0 likes · 5 min read

This 103‑page survey of over 300 recent papers organizes large language model reasoning into a periodic‑table framework, explains where reasoning emerges, categorizes 36 method families across six dimensions, critiques accuracy‑only evaluation, and outlines key open challenges such as fidelity, robustness, calibration, generalization, efficiency, and safety.
AI safetyLLM reasoningMeta-reasoning
0 likes · 13 min read

TaH, a selective latent‑iteration method for small language models, identifies and avoids unnecessary token‑level loops, cutting about 93% of extra iterations while delivering a stable 3.0%‑6.8% accuracy gain across nine math, QA, and code benchmarks.
LLM reasoningLooped TransformerTaH
0 likes · 14 min read

The article explains the limitations of large language models on multi‑step reasoning, real‑time information retrieval, and planning, then introduces the ReAct (Reasoning + Acting) framework and its Reflexion extension, detailing their mechanisms, examples, performance gains, practical applications, and future research directions.
Agentic AILLM reasoningPrompt engineering
0 likes · 16 min read

The article presents the OnePiece framework, which injects LLM‑style context engineering and latent reasoning into item‑ID based search‑and‑recommendation models, details the design choices, training tricks, attention analysis, and reports online gains of around 1% GMV and ad revenue, offering a thorough technical dissection of generative recommendation in industrial settings.
LLM reasoningOnePieceTransformer
0 likes · 31 min read

DeepSeek‑R1 demonstrates that large‑scale reinforcement learning, especially with the novel Group Relative Policy Optimization and a rule‑based reward scheme, can markedly boost reasoning in LLMs without heavy supervised fine‑tuning, while a brief cold‑start SFT phase, two‑stage alignment, and knowledge distillation further improve performance and efficiency, despite remaining challenges such as language mixing.
DeepSeek-R1GRPOLLM reasoning
0 likes · 21 min read


ArchitectFeb 6, 2025 · Artificial Intelligence
The article reviews DeepSeek‑R1, detailing its reinforcement‑learning‑based training pipeline that uses minimal supervised data, cold‑start fine‑tuning, multi‑stage RL, rejection‑sampling SFT, and distillation to achieve reasoning performance comparable to OpenAI‑o1‑1217, while also discussing successful contributions and failed experiments.
AI researchDeepSeek-R1LLM reasoning
0 likes · 11 min read

Examining OpenAI’s o1 model, this article explores its self‑correction capability by linking test‑time scaling, MCTS‑style reasoning, and DeepMind’s SCoRe reinforcement‑learning framework, illustrating step‑by‑step demos, hypothesizing internal judgment mechanisms, and proposing training pipelines that combine self‑generated data with post‑training RL.
LLM reasoningMCTSOpenAI
0 likes · 12 min read
