๋ณธ๋ฌธ์œผ๋กœ ๊ฑด๋„ˆ๋›ฐ๊ธฐ

Topic: #self-improvement

A collection of 24 posts about self-improvement.

NTU ํ•˜ํŽธ 51โ€“55์ชฝ. ํ† ํฐ์ด TTT ์ธต์˜ ๊ฐ€์ค‘์น˜๋ฅผ ๋ฐ”๊พธ๋Š” ๊ณผ์ •, Titans์™€ Nested Learning์˜ ๋ฉ”๋ชจ๋ฆฌ, ๋ฌธ๋งฅ ์† ์ง€์‹œ์„œ๊ฐ€ ๋‹ต์„ ๋ฐ”๊พธ๋Š” ๊ณผ์ •์„ ์ผ๊ณฑ ์˜ˆ์ œ๋กœ ๊ณ„์‚ฐํ•œ๋‹ค.

15 min read
#self-improvement#test-time-training#memory
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU ํ•˜ํŽธ 46โ€“50์ชฝ์„ ๋”ฐ๋ผ SEAL์˜ ์ž๊ธฐ ํŽธ์ง‘ ๋ณด์ƒ, ๋‘ ํ•™์Šต ์‹œ๊ฐ„ ์ฒ™๋„, ํ•™์Šต๋œ ์ตœ์ ํ™”๊ธฐ์™€ MAML์„ ์‚ดํŽด๋ณธ๋‹ค. ์‹ ํ’€์ด ์ผ๊ณฑ ๊ฐœ๋กœ ๋ชจ๋ธ๊ณผ ํ•™์Šต ๊ทœ์น™์˜ ๊ฐฑ์‹ ์„ ๊ตฌ๋ถ„ํ•œ๋‹ค.

15 min read
#self-improvement#ai-agents#prompt-optimization
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU ํ•˜ํŽธ 41โ€“45์ชฝ์˜ ์‹ ํ’€์ด 7๊ฐœ๋กœ ๊ฒฝ์‚ฌํ•˜๊ฐ•๋ฒ•, PostTrainBench, autoresearch, AlphaEvolve์™€ ShinkaEvolve๋ฅผ ์ฝ๋Š”๋‹ค. ๊ฐ€์ค‘์น˜ ํ•œ ๋ฒˆ์˜ ๊ฐฑ์‹ ์—์„œ ํ•™์Šต ๊ทœ์น™ ์ž์ฒด์˜ ํ‰๊ฐ€๊นŒ์ง€ ๊ณ„์‚ฐํ•œ๋‹ค.

14 min read
#self-improvement#ai-agents#prompt-optimization
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU ํ•˜ํŽธ 36โ€“40์ชฝ์„ ๋”ฐ๋ผ ํ•˜๋„ค์Šค๊ฐ€ ๋‹ค์Œ ํŽธ์ง‘์ž๊ฐ€ ๋˜๋Š” ๊ณผ์ •์„ ์ฝ๋Š”๋‹ค. ์‹ ํ’€์ด ์ผ๊ณฑ ๊ฐœ๋กœ Hyperagents์˜ ๋ถ€๋ชจ ์„ ํƒ ์‹คํ—˜๊ณผ Learning to Self-Evolve์˜ ๋ฌธ๋งฅ ํŽธ์ง‘ ๋ณด์ƒ์„ ๊ณ„์‚ฐํ•œ๋‹ค.

15 min read
#self-improvement#ai-agents#prompt-optimization
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU ํ•˜ํŽธ 31โ€“35์ชฝ์„ ์ฝ์œผ๋ฉฐ ํ•˜๋„ค์Šค๋ฅผ ๋ฐ”๊ฟ€ ๋•Œ ์ƒ๊ธฐ๋Š” ๋ง๊ฐ๊ณผ ๊ฐœ์„ ๊ธฐ์˜ ์—ญํ• ์„ ์‚ดํŽด๋ณธ๋‹ค. ๊ฐ€์ค‘์น˜๋ฅผ ๊ณ ์ •ํ•ด๋„ ์˜ˆ์ „ ๋Šฅ๋ ฅ์ด ๋–จ์–ด์ง€๋Š” ์ด์œ , ์„ฑ๊ณต๋ฅ ๊ณผ ์œ ์ง€์œจ์˜ ๋ถ„๋ชจ, DGM์˜ ํŽธ์ง‘์ž ๊ณ ์ • ์‹คํ—˜์„ ์—ฌ์„ฏ ๊ณ„์‚ฐ ์˜ˆ์ œ๋กœ ํ‘ผ๋‹ค.

13 min read
#self-improvement#ai-agents#prompt-optimization
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU ํ•˜ํŽธ 26โ€“30์ชฝ์„ ์ฝ์œผ๋ฉฐ ์ž…๋ ฅ๋งˆ๋‹ค ๋ชจ๋ธ์„ ์ดˆ๊ธฐํ™”ํ•˜๋Š” ๋ฐฉ๋ฒ•๊ณผ ์ ์‘ ๊ฒฐ๊ณผ๋ฅผ ์ด์–ด๋ฐ›๋Š” ๋ฐฉ๋ฒ•์„ ๋น„๊ตํ•œ๋‹ค. ๋จธ์‹ ๋น„์ „์˜ ํšŒ์ „ ์˜ˆ์ธก, ์†Œ์Œ ์† ์Œ์„ฑ์ธ์‹ ์‹คํ—˜, ๊ฒฝํ—˜ ์žฌ์ƒ์œผ๋กœ ์ƒˆ ํ™˜๊ฒฝ ์ ์‘๊ณผ ๊ธฐ์กด ๋Šฅ๋ ฅ ๋ณด์กด์„ ์„ค๋ช…ํ•œ๋‹ค.

15 min read
#self-improvement#test-time-training#test-time-adaptation
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU 2026 ํ•˜ํŽธ 21โ€“25์ชฝ์„ ํ•œ ์žฅ์”ฉ ์ฝ๋Š”๋‹ค. ExpRAG์˜ ๊ฒฝํ—˜ ๊ฒ€์ƒ‰๊ณผ Fine-tuning, BetterTogether์˜ ๊ต๋Œ€ ์ตœ์ ํ™”, E-SPL์˜ ํ”„๋กฌํ”„ํŠธ ์ง„ํ™”์™€ ๊ฐ•ํ™”ํ•™์Šต, ๋ชฉํ‘œ ๋ณ€ํ™” ์‹œ ์žฌํ‰๊ฐ€๋ฅผ ๊ณต์‹ ๋…ผ๋ฌธ ๋ฐ Omar Khattab ๊ฐ•์˜๋กœ ํ•ด์„คํ•œ๋‹ค.

16 min read
#self-improvement#ai-agents#prompt-optimization
๐Ÿ‡ฐ๐Ÿ‡ท ko

NTU ์ž๊ฐ€๊ฐœ์„  AI ํ•˜ํŽธ 16โ€“20์ชฝ์„ ์›๋ณธ ์Šฌ๋ผ์ด๋“œ ํ•œ ์žฅ์”ฉ ์ฝ๋Š”๋‹ค. DGM ์•„์นด์ด๋ธŒ์™€ ํƒ์ƒ‰ ๊ณ„๋ณด, DSPy์˜ ์„ ์–ธํ˜• ํ”„๋กœ๊ทธ๋žจ๊ณผ ์ž๋™ ์ตœ์ ํ™”, ํ•˜๋„ค์Šคยท๋ชจ๋ธ ํŒŒ๋ผ๋ฏธํ„ฐ์˜ ๊ฒฝ๊ณ„, ๊ฒฝํ—˜ ๊ฒ€์ƒ‰๊ณผ fine-tuning์„ ๋…ผ๋ฌธ๊ณผ ์™ธ๋ถ€ ๊ฐ•์˜ ์ž๋ฃŒ๋กœ ๋ณด๊ฐ•ํ•œ๋‹ค.

14 min read
#self-improvement#ai-agents#dspy
๐Ÿ‡ฐ๐Ÿ‡ท ko

ํ•˜ํŽธ 11โ€“15์ชฝ์˜ OPROยทGEPAยทALMAยทDarwin Gรถdel Machine์„ ํ•œ ์žฅ์”ฉ ์ฝ๋Š”๋‹ค. CMUยทStanford ๋ฐœํ‘œ ์Šฌ๋ผ์ด๋“œ์™€ GEPA ์ €์ž์˜ ์˜์ƒ ์„ค๋ช…์„ ์—ฐ๊ฒฐํ•ด ํ”„๋กฌํ”„ํŠธ ํƒ์ƒ‰, ๋ฌธํ•ญ๋ณ„ ํ›„๋ณด ์„ ํƒ, ๋ฉ”๋ชจ๋ฆฌ ์„ค๊ณ„์™€ ์ฝ”๋“œ ์ž๊ธฐ์ˆ˜์ •์˜ ์›๋ฆฌยท๊ณ„์‚ฐ ์˜ˆยทํ‰๊ฐ€ ์กฐ๊ฑด์„ ์„ค๋ช…ํ•œ๋‹ค.

24 min read
#self-improvement#prompt-optimization#ai-agents
๐Ÿ‡ฐ๐Ÿ‡ท ko

ํ•˜ํŽธ 6โ€“10์ชฝ์—์„œ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜์™€ ํ•˜๋„ค์Šค๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ณ , ํ”„๋กฌํ”„ํŠธยท๋„๊ตฌยท์‹คํ–‰ ํ๋ฆ„์„ ์ˆ˜์ •ํ•˜๋Š” ์›๋ฆฌ๋ฅผ ์‚ดํŽด๋ณธ๋‹ค. CMUยทNAIRRยทMIT ๊ฐ•์˜์™€ SWE-agentยทDSPyยทAFlowยทฯ„-bench๋ฅผ ์—ฐ๊ฒฐํ•ด ๊ธฐ์–ต ๊ด€๋ฆฌ, ํ›„๋ณด ์ƒ์„ฑ, ๊ณต์ •ํ•œ ๋น„๊ต, ๋น„์šฉ๊ณผ ๋ฐ˜๋ณต ์‹ ๋ขฐ์„ฑ์„ ์„ค๋ช…ํ•œ๋‹ค.

24 min read
#self-improvement#ai-agents#harness
๐Ÿ‡ฐ๐Ÿ‡ท ko

ํ•˜ํŽธ 1โ€“5์ชฝ์˜ ํ•™์Šต ํ๋ฆ„์„ ๋”ฐ๋ผ ํ›ˆ๋ จ ์†์‹ค๊ณผ ์‹ค์ œ ํ‰๊ฐ€๋ฅผ ๊ตฌ๋ถ„ํ•˜๊ณ , Absolute ZeroยทR-ZeroยทSPICEยทR-Few๊ฐ€ ์ถœ์ œ์žยทํ’€์ด์žยท๊ฒ€์ฆ์ž๋กœ ์ž๊ธฐํ•™์Šต์„ ๊ตฌ์„ฑํ•˜๋Š” ์›๋ฆฌ์™€ ํ•œ๊ณ„๋ฅผ ์‚ดํŽด๋ณธ๋‹ค.

26 min read
#self-improvement#self-play#reinforcement-learning
๐Ÿ‡ฐ๐Ÿ‡ท ko

AI ์—์ด์ „ํŠธ๋ฅผ ๋ชจ๋ธ ๊ฐ€์ค‘์น˜, ํ•˜๋„ค์Šค, ์‹คํ–‰ ์ƒํƒœ๋กœ ๋‚˜๋ˆ„๊ณ  ReActยทGenerative AgentsยทReflexionยทSWE-agent์˜ ๊ตฌ์กฐ๋ฅผ ํ†ตํ•ด ๋„๊ตฌ์™€ ๋ฉ”๋ชจ๋ฆฌ๊ฐ€ ํ–‰๋™์„ ๋ฐ”๊พธ๋Š” ์›๋ฆฌ์™€ ๊ณต์ •ํ•œ ๊ฐœ์„  ํ‰๊ฐ€๋ฅผ ์‚ดํŽด๋ณธ๋‹ค.

14 min read
#self-improvement#ai-agents#harness
๐Ÿ‡ฐ๐Ÿ‡ท ko

Weak-to-Strong Generalization์˜ ํ•™์Šต ๊ตฌ์กฐ์™€ ์„ฑ๋Šฅ ๊ฒฉ์ฐจ ํšŒ๋ณต๋ฅ  PGR์„ ์„ค๋ช…ํ•˜๊ณ , OpenAI์˜ ๋ณด์กฐ ํ™•์‹  ์†์‹ค๊ณผ Anthropic ์ž๋™ ์ •๋ ฌ ์—ฐ๊ตฌ์ž์˜ ์„ฑ๊ณผยท์ „์ด ํ•œ๊ณ„๋ฅผ ์‚ดํŽด๋ณธ๋‹ค.

17 min read
#self-improvement#weak-to-strong#alignment
๐Ÿ‡ฐ๐Ÿ‡ท ko

PostTrainBench์™€ FT-Dojo๋กœ ํ›„ํ•™์Šต ์ž๋™ํ™”๋ฅผ ์‚ดํŽด๋ณธ๋‹ค. ํ•™์Šต ์—์ด์ „ํŠธ์™€ ๋Œ€์ƒ ๋ชจ๋ธ์„ ๊ตฌ๋ถ„ํ•˜๊ณ , ์‹œ๊ฐ„ ์˜ˆ์‚ฐยท๊ฐ€์ค‘ ํ‰๊ฐ€ ์ ์ˆ˜ยท๋ฐ์ดํ„ฐ ์˜ค์—ผ์„ ํ•จ๊ป˜ ์ฝ๋Š”๋‹ค.

16 min read
#self-improvement#post-training#llm-agents
๐Ÿ‡ฐ๐Ÿ‡ท ko

๋ฌธ์ œ ์ƒ์„ฑ์ž์™€ ํ’€์ด์ž์˜ ์ž๊ธฐ๋Œ€๊ฒฐ, R-Zero์˜ ์„ฑ๋Šฅ ์ •์ฒด, Absolute Zero์˜ ์‹คํ–‰ ๊ฒ€์ฆ, SPICE์™€ R-Few์˜ ์™ธ๋ถ€ ์ •๋ณด ํ™œ์šฉ์„ ์‚ดํŽด๋ณธ๋‹ค.

15 min read
#self-improvement#self-play#reinforcement-learning
๐Ÿ‡ฐ๐Ÿ‡ท ko

์—”ํŠธ๋กœํ”ผ๊ฐ€ ๋‚ฎ์€ ๊ฒฝ๋กœ๋ฅผ ์„ ํƒํ•˜๋Š” ์ •์ฑ… ๊ฒฝ์‚ฌ, ์Œ์„ฑ์ธ์‹ WER ๋น„๊ต, AI ํ‰๊ฐ€์™€ ํ•™์Šต ๋ฌธ์ œ ์ƒ์„ฑ์˜ ์—ญํ• ์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.

18 min read
#self-improvement#entropy-minimization#policy-gradient
๐Ÿ‡ฐ๐Ÿ‡ท ko

๋ฌธ์žฅ ์—”ํŠธ๋กœํ”ผ์˜ ํ† ํฐ๋ณ„ ์ถ”์ •๊ณผ ์ •์ฑ… ๊ฒฝ์‚ฌ๋ฅผ ์—ฐ๊ฒฐํ•˜๊ณ , ์ƒ˜ํ”Œ์„ ๊ณ ์ •ํ•˜๋ฉด ๋น ์ง€๋Š” ํ•ญ์„ ํ™•๋ฅ ๋‚˜๋ฌด์™€ ์ž๋™๋ฏธ๋ถ„ ์ฝ”๋“œ๋กœ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.

15 min read
#self-improvement#entropy-minimization#policy-gradient
๐Ÿ‡ฐ๐Ÿ‡ท ko

์ž๊ธฐํ‰๊ฐ€ ์ ์ˆ˜, ๋‹ค์ˆ˜๊ฒฐ, ์—”ํŠธ๋กœํ”ผ๋Š” ๋ฌด์—‡์ด ๋‹ค๋ฅผ๊นŒ? TentยทSUTAยทLLM ์—”ํŠธ๋กœํ”ผ ์ตœ์†Œํ™”์˜ ํ•™์Šต ์‹ ํ˜ธ์™€ ์ ์šฉ ์กฐ๊ฑด์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.

19 min read
#self-improvement#self-evaluation#entropy-minimization
๐Ÿ‡ฐ๐Ÿ‡ท ko

์ž…๋ ฅ ํผํ”Œ๋ ‰์„œํ‹ฐ, ๋น„์ง€๋„ RLVR์˜ ์ดˆ๊ธฐ ํ–ฅ์ƒ๊ณผ ๋ถ•๊ดด, ํ…Œ์ŠคํŠธ ์‹œ์  ํ•™์Šต์„ ์‚ดํŽด๋ณด๊ณ  ์ดˆ๊ธฐํ™”ํ˜•ยท๋ˆ„์ ํ˜• ์ ์‘์˜ ์ฐจ์ด๋ฅผ ๊ณ„์‚ฐํ•ฉ๋‹ˆ๋‹ค.

16 min read
#self-improvement#unsupervised-rlvr#test-time-training
๐Ÿ‡ฐ๐Ÿ‡ท ko

์ž๊ฐ€๊ฐœ์„  AI์—์„œ ๋ฐ”๋€Œ๋Š” ๋Œ€์ƒ๊ณผ ํ”ผ๋“œ๋ฐฑ์˜ ์—ญํ• ์„ ์‚ดํŽด๋ณด๊ณ , ์žฌ๊ท€์  ์ž๊ธฐ๊ฐœ์„ ์˜ ์ „๋ง์„ ์‹ค์ œ ํ‰๊ฐ€ยท์ผ๋ฐ˜ํ™”์˜ ๊ทผ๊ฑฐ์™€ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค.

18 min read
#self-improvement#self-correction#reinforcement-learning
๐Ÿ‡ฐ๐Ÿ‡ท ko

ํ•™์Šต ๋ชฉํ‘œ์™€ ์†์‹ค ํ•จ์ˆ˜์—์„œ ์ถœ๋ฐœํ•ด ์ง€์‹ ์ฆ๋ฅ˜์™€ ์ž๊ธฐ ์ƒ์„ฑ ์˜์‚ฌ ์ •๋‹ต์„ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ๊ฐ™์€ ๋ชจ๋ธ์˜ ์ถœ๋ ฅ์ด ์–ธ์ œ ์ƒˆ๋กœ์šด ํ•™์Šต ์‹ ํ˜ธ๊ฐ€ ๋˜๋Š”์ง€ ๊ณ„์‚ฐ์œผ๋กœ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.

15 min read
#self-improvement#knowledge-distillation#self-training
๐Ÿ‡ฐ๐Ÿ‡ท ko

๋‹ต์„ ๊ณ ์น˜๋Š” ์ž๊ธฐ์ˆ˜์ •๊ณผ ํŒŒ๋ผ๋ฏธํ„ฐ๋ฅผ ๋ฐ”๊พธ๋Š” ํ•™์Šต์„ ๊ตฌ๋ถ„ํ•ฉ๋‹ˆ๋‹ค. ์ˆ˜์ • ๋‹ต์˜ ์ง€๋„ํ•™์Šต, ์ˆ˜์ • ํ–‰๋™์˜ ๊ฐ•ํ™”ํ•™์Šต, ๋ณด์ƒ์— ๋”ฐ๋ฅธ ์—…๋ฐ์ดํŠธ๋ฅผ ๊ณ„์‚ฐ ์˜ˆ์‹œ์™€ ํ•จ๊ป˜ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.

19 min read
#self-improvement#self-correction#reinforcement-learning
๐Ÿ‡ฐ๐Ÿ‡ท ko

ํฌ์†Œ ๋ณด์ƒ์˜ ์–ด๋ ค์›€๊ณผ ์ค‘๊ฐ„ ๋ณด์ƒ์ด ๋ชฉํ‘œ๋ฅผ ๋ฐ”๊พธ๋Š” ๋ฐ˜๋ก€๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค. ์ž ์žฌ ํ•จ์ˆ˜ ๊ธฐ๋ฐ˜ ๋ณด์ƒ์˜ ์กฐ๊ฑด, LLM ๋ณด์ƒ ์ฝ”๋“œ ์ƒ์„ฑ๊ณผ ์ •์ฑ… ํ‰๊ฐ€๋ฅผ ์—ฐ๊ฒฐํ•ฉ๋‹ˆ๋‹ค.

17 min read
#self-improvement#reinforcement-learning#reward-shaping
๐Ÿ‡ฐ๐Ÿ‡ท ko

EurekaยทREvolveยทRF-Agent์˜ ๋ณด์ƒ ํƒ์ƒ‰, ๋„ํŒŒ๋ฏผ์˜ ๋ณด์ƒ ์˜ˆ์ธก ์˜ค์ฐจ, ์„ ํ˜ธ ๊ธฐ๋ฐ˜ ๋ณด์ƒ ๋ชจ๋ธ๊ณผ RLHFยทRLAIF์˜ ์ฐจ์ด๋ฅผ ์‚ดํŽด๋ด…๋‹ˆ๋‹ค.

14 min read
#self-improvement#reinforcement-learning#reward-shaping
๐Ÿ‡ฐ๐Ÿ‡ท ko