TAHI: Test-Time Adaptive Agents That Fit Individual Experts in Tens of Tasks
AkariAsai · x · 2026-10-05
Researcher Zhiruo Wang presents TAHI, a test-time adaptive agent framework built through human-agent interaction, ahead of a CoLM2026 workshop talk. Key points:
- Test-time adaptation via both context (memory, skills) and weight training
- Efficient personalization: adapts to an individual's expertise within tens of tasks
- Rubrics for "non-verifiable" tasks where "pretty good" isn't good enough for professionals staking their reputation
- Analysis of shared community guidelines vs. personalized tacit expertise
Core thesis: agents trained on population-scale data can do a lot, but professionals need far more reliability. The author is also on the academic job market.
More from coding & agent
- He turned an old laptop and Claude-written scripts into a 24/7 video line to his mom — spiritdouble · 2026-10-05
- Opus 5.5 puts Minecraft inside Slime Rancher — and it actually works — Angaisb_ · 2026-10-05
- GitHarness paper: managing agent memory like a Git repo to handle mid-task requirement shifts — rohanpaul_ai · 2026-10-05
- GitHarness: agent memory as a Git repo cuts tokens 73.6% in one coding setup — rohanpaul_ai · 2026-10-05
- Stripping antirez's ds4 to 45k lines makes Qwen3.8 Flash Next ~10% faster, bit-exact — Chida82 · 2026-10-05
- Same prompt showdown: Claude vs ChatGPT vs Figma Make — Tegadesigns · 2026-10-05