--- title: How Production LLMs Reason Better At Inference Time description: An explainer covering eight inference-time reasoning techniques for LLMs, split between parallel scaling (chain of thought, majority voting, best-of-N) and... canonical: https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: How Production LLMs Reason Better At Inference Time | daily.dev og:description: An explainer covering eight inference-time reasoning techniques for LLMs, split between parallel scaling (chain of thought, majority voting, best-of-N) and... og:url: https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue og:image: https://api.daily.dev/og/posts/WeygdoSuE.png og:image:alt: How Production LLMs Reason Better At Inference Time og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # How Production LLMs Reason Better At Inference Time **[Daily Dose of Data Science \| Avi Chawla \| Substack](https://daily.dev/sources/dailydoseofds)** · 8 min read · 1 upvotes · 0 comments ## Summary An explainer covering eight inference-time reasoning techniques for LLMs, split between parallel scaling (chain of thought, majority voting, best-of-N) and sequential scaling (extended thinking, self-refinement, Tree of Thought, beam search with process reward models, and MCTS). It cites research findings such as R1-Zero's jump from 15.6% to 71.0% on AIME 2024 through pure RL, Anthropic's inverse scaling work showing longer reasoning traces can hurt accuracy, and DeepSeek's decision to drop MCTS and PRMs in favor of rule-based rewards for R1. It opens with an aside on why smarter Claude models burn more tokens on backend tasks, contrasting Supabase's dashboard-oriented MCP interface (10.4M tokens, 10 manual interventions on a test build) against InsForge, an open-source agent-first backend that returned full topology in ~500 tokens and completed the same build in 3.7M tokens with zero errors. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why does using a smarter Claude model in Claude Code increase token usage instead of decreasing it on backend tasks? A smarter Claude model uses 54% more tokens on backend tasks, per MCPMark V2 benchmarks across 21 backend tasks, because the extra capability goes into trying harder to fill context gaps left by backends not designed for agents, such as Supabase returning entire auth documentation instead of the specific OAuth setup info requested, forcing more discovery queries and retries. _daily.dev surfaces practical findings like this for teams tuning agent-driven backend workflows and token costs._ ### What is the difference between best-of-N and beam search with a process reward model for LLM inference-time reasoning? Best-of-N generates N complete answers and scores each with a reward model at the end, while beam search with a process reward model keeps K partial solutions alive and scores every step as it is written. Beam search outperforms best-of-N at small compute budgets but falls behind as budgets grow, since over-optimized runs produce repetitive or collapsed reasoning steps. _developers comparing inference-time scaling strategies can track technique tradeoffs like these on daily.dev._ ### Why did DeepSeek drop Monte Carlo Tree Search (MCTS) and process reward models (PRMs) when building R1? DeepSeek dropped MCTS because the number of possible next tokens vastly exceeds the legal moves in a board game, making it too hard to train a value model good enough to guide the search. PRMs were dropped because step correctness is hard to define, labeling does not scale, and the policy learns to exploit the scorer, so DeepSeek used rule-based rewards instead. _teams evaluating reasoning-model training approaches can follow reports like this on daily.dev._ ## Similar posts on daily.dev - [Categories of Inference-Time Scaling for Improved LLM Reasoning](https://daily.dev/posts/categories-of-inference-time-scaling-for-improved-llm-reasoning-bguxcwmhk) · Sebastian Raschka · 22 upvotes · 0 comments - [Don’t be fooled—LLMs don’t reason](https://daily.dev/posts/don-t-be-fooled-llms-don-t-reason-4waw8d0j4) · MIT Technology Review · 0 upvotes · 1 comments - [New method could increase LLM training efficiency](https://daily.dev/posts/new-method-could-increase-llm-training-efficiency-oclgiwdge) · MIT News · 0 upvotes · 0 comments - [Do engineers still need to understand how LLMs work?](https://daily.dev/posts/do-engineers-still-need-to-understand-how-llms-work--yqrlcrtga) · Tech World With Milan · 1 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"How Production LLMs Reason Better At Inference Time","url":"https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue"},"datePublished":"2026-08-14T20:48:02.949Z","dateModified":"2026-09-14T06:54:25.163Z","description":"An explainer covering eight inference-time reasoning techniques for LLMs, split between parallel scaling (chain of thought, majority voting, best-of-N) and...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8eed2d805a9df17e612602e90e5c0fb0?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8eed2d805a9df17e612602e90e5c0fb0?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Daily Dose of Data Science | Avi Chawla | Substack","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Daily Dose of Data Science | Avi Chawla | Substack","logo":"https://media.daily.dev/image/upload/s--4IHQgTOw--/f_auto/v1710503712/logos/dailydoseofds","url":"https://daily.dev/sources/dailydoseofds"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,claude,reinforcement-learning","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Daily Dose of Data Science | Avi Chawla | Substack","item":"https://daily.dev/sources/dailydoseofds"},{"@type":"ListItem","position":3,"name":"How Production LLMs Reason Better At Inference Time"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/how-production-llms-reason-better-at-inference-time-weygdosue#faq","mainEntity":[{"@type":"Question","name":"Why does using a smarter Claude model in Claude Code increase token usage instead of decreasing it on backend tasks?","acceptedAnswer":{"@type":"Answer","text":"A smarter Claude model uses 54% more tokens on backend tasks, per MCPMark V2 benchmarks across 21 backend tasks, because the extra capability goes into trying harder to fill context gaps left by backends not designed for agents, such as Supabase returning entire auth documentation instead of the specific OAuth setup info requested, forcing more discovery queries and retries. daily.dev surfaces practical findings like this for teams tuning agent-driven backend workflows and token costs."}},{"@type":"Question","name":"What is the difference between best-of-N and beam search with a process reward model for LLM inference-time reasoning?","acceptedAnswer":{"@type":"Answer","text":"Best-of-N generates N complete answers and scores each with a reward model at the end, while beam search with a process reward model keeps K partial solutions alive and scores every step as it is written. Beam search outperforms best-of-N at small compute budgets but falls behind as budgets grow, since over-optimized runs produce repetitive or collapsed reasoning steps. developers comparing inference-time scaling strategies can track technique tradeoffs like these on daily.dev."}},{"@type":"Question","name":"Why did DeepSeek drop Monte Carlo Tree Search (MCTS) and process reward models (PRMs) when building R1?","acceptedAnswer":{"@type":"Answer","text":"DeepSeek dropped MCTS because the number of possible next tokens vastly exceeds the legal moves in a board game, making it too hard to train a value model good enough to guide the search. PRMs were dropped because step correctness is hard to define, labeling does not scale, and the policy learns to exploit the scorer, so DeepSeek used rule-based rewards instead. teams evaluating reasoning-model training approaches can follow reports like this on daily.dev."}}]} ```