https://cognaptus.com/tags/llm-training/ Recent content in LLM Training on Cognaptus Hugo -- 0.145.0 en-us Wed, 03 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-03-synthetic-and-sensibility-why-more-data-needs-a-control-stack/ Wed, 03 Jun 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-06-03-synthetic-and-sensibility-why-more-data-needs-a-control-stack/ Synthetic data becomes useful only when it is verified, diversified, matched to the student model, and audited for downstream transfer. https://cognaptus.com/blog/2026-05-26-adamw-and-the-cost-of-being-reasonable-choosing-llm-optimizers-without-leaderboard-theater/ Tue, 26 May 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-05-26-adamw-and-the-cost-of-being-reasonable-choosing-llm-optimizers-without-leaderboard-theater/ A business-facing reading of why LLM optimizer choice is less about replacing AdamW and more about trading memory, stability, wall-clock time, and hardware fit. https://cognaptus.com/blog/2026-04-08-when-data-decides-what-matters-the-quiet-economics-of-llm-data-selection/ Wed, 08 Apr 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-04-08-when-data-decides-what-matters-the-quiet-economics-of-llm-data-selection/ A clearer look at why dynamic data weighting may matter less as a magic shortcut than as a new control layer for LLM training economics. https://cognaptus.com/blog/2026-03-16-when-right-meets-wrong-teaching-llms-by-letting-their-mistakes-talk/ Mon, 16 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-16-when-right-meets-wrong-teaching-llms-by-letting-their-mistakes-talk/ A mechanism-first reading of BiCC and RCC, showing how successful and failed reasoning traces can improve GRPO-style training without adding inference-time overhead. https://cognaptus.com/blog/2026-03-12-mirror-mirror-on-the-agent-teaching-llms-to-judge-their-own-actions/ Thu, 12 Mar 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-03-12-mirror-mirror-on-the-agent-teaching-llms-to-judge-their-own-actions/ A mechanism-first reading of Agentic Critical Training and why teaching agents to compare actions may matter more than teaching them to explain themselves. https://cognaptus.com/blog/2026-02-24-resyn-the-rise-of-the-verifier-when-solving-is-hard-but-checking-is-easy/ Tue, 24 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-24-resyn-the-rise-of-the-verifier-when-solving-is-hard-but-checking-is-easy/ ReSyn shows why scalable reasoning training may depend less on generating more answers and more on building synthetic environments where correctness can be checked reliably. https://cognaptus.com/blog/2026-02-21-from-static-models-to-living-systems-when-ai-stops-predicting-and-starts-adapting/ Sat, 21 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-21-from-static-models-to-living-systems-when-ai-stops-predicting-and-starts-adapting/ A business-focused reading of dynamic bi-level data weighting, and why the next training advantage may come from adaptive data utilization rather than simply larger datasets. https://cognaptus.com/blog/2026-02-13-breaking-things-on-purpose-how-cligym-teaches-ai-to-fix-the-real-world/ Fri, 13 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-13-breaking-things-on-purpose-how-cligym-teaches-ai-to-fix-the-real-world/ A mechanism-first reading of CLI-Gym, a pipeline that turns working Dockerized repositories into scalable environment-repair tasks for stronger coding agents. https://cognaptus.com/blog/2026-02-09-agents-need-worlds-not-prompts-inside-scaleenvs-synthetic-environment-revolution/ Mon, 09 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-09-agents-need-worlds-not-prompts-inside-scaleenvs-synthetic-environment-revolution/ ScaleEnv shows why serious tool-use agents need executable, stateful, verifiable training worlds—not just better prompts or prettier tool-call examples. https://cognaptus.com/blog/2026-02-08-freeze-now-learn-faster-when-parameter-freezing-meets-pipeline-reality/ Sun, 08 Feb 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-02-08-freeze-now-learn-faster-when-parameter-freezing-meets-pipeline-reality/ TimelyFreeze shows that parameter freezing only becomes a real training-speed lever when it is aligned with the pipeline schedule’s wall-clock bottlenecks. https://cognaptus.com/blog/2026-01-17-matchtir-stop-paying-every-token-the-same-salary/ Sat, 17 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-17-matchtir-stop-paying-every-token-the-same-salary/ MatchTIR shows why multi-turn tool agents need fine-grained credit assignment, not just bigger models or louder final-answer rewards. https://cognaptus.com/blog/2026-01-03-when-models-start-to-forget-the-hidden-cost-of-training-llms-too-well/ Sat, 03 Jan 2026 00:00:00 +0000 https://cognaptus.com/blog/2026-01-03-when-models-start-to-forget-the-hidden-cost-of-training-llms-too-well/ A practical reading of why LLM memorization becomes hard to remove once training entangles recall with general capability. https://cognaptus.com/blog/2025-12-31-browsing-without-the-bloat-teaching-agents-to-think-before-they-scroll/ Wed, 31 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-31-browsing-without-the-bloat-teaching-agents-to-think-before-they-scroll/ NestBrowse shows that better browser agents may depend less on larger models or longer contexts than on controlling which information reaches the reasoning loop. https://cognaptus.com/blog/2025-12-09-no-prompt-left-behind-how-shopees-compassmax-reinvents-rl-for-giant-moe-models/ Tue, 09 Dec 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-12-09-no-prompt-left-behind-how-shopees-compassmax-reinvents-rl-for-giant-moe-models/ Shopee’s CompassMax-V3-Thinking paper shows that scaling RL for giant MoE models is less about buying more rollouts and more about making every rollout produce usable learning signal. https://cognaptus.com/blog/2025-07-23-weight-watchers-for-llms-dynamic-dieting-beats-static-selection/ Wed, 23 Jul 2025 00:00:00 +0000 https://cognaptus.com/blog/2025-07-23-weight-watchers-for-llms-dynamic-dieting-beats-static-selection/ A mechanism-first reading of why dynamic data weighting may matter more than static corpus selection for efficient LLM pretraining.