Training stages as additive task vectors: base + midtrain + RLVR surprisingly just adds up
voooooogel · x · 2026-10-29
A technical thread on decomposing multi-stage training (base → midtrain → RLVR) into additive task vectors. The author notes that while stages are theoretically entangled—e.g. whether midtraining suppresses reward hacking shapes the (rlvr - midtrain) vector—in practice these vectors surprisingly add linearly and behave as expected, echoing the "unreasonable effectiveness of model merging." This enables grafting a new midtraining stage without a full retrain: just compute (rlvr + (newmidtrain - base)), and it works surprisingly well.
Related event: Multi-stage LLM training decomposes into additive task vectors(2 posts)→
More from Research
- Persona prompt pins models at ~35% on SIMMER benchmark regardless of Claude or DeepSeek — amplifiedamp · 2026-10-29
- Alfred Menezes releases gentle introduction to lattice-based crypto covering Kyber and Dilithium — matthew_d_green · 2026-10-29
- Google DeepMind, NUH Singapore Start Trusted Tester Trials of AI Co-Clinician — alan_karthi · 2026-10-29
- NPCI Releases TauIndianBankBench for Tool-Using Banking Agents — NPCI · 2026-10-29
- Axiom's agents are formalizing CFSG in 43M lines of code, possibly the largest open-source project ever — SucceededMind · 2026-10-29
- MIT's 'AI telepathy' KV-cache transfer beats text by 40% accuracy at 4x speed, across model vendors — LucaAmb · 2026-10-29