Agent success decomposes into syntax × argument fidelity, plus retry recovery. Long trajectories blur outcome credit to ~1/√steps — hence process rewards.
Post-training extends beyond text: interleaved image/audio tokens need their own alignment data, and agentic models must not just answer but act — emitting well-formed tool calls with faithful arguments. Both extend the SFT/RL toolkit with new reward surfaces.
Token Mix
An image costs hundreds-to-thousands of tokens via a vision encoder. Token budgeting across modalities decides context economics and which modality dominates gradients.
Tool-Call Success
Success decomposes: syntax valid × arguments faithful × execution succeeds. Training on execution traces (with retries) teaches repair behavior.
Trajectory Credit
A 30-step agent run has one outcome signal. Per-step credit blurs as 1/√steps — hence step-level verifiers, process rewards, and short-horizon curricula.
95% syntax × 90% argument fidelity = 85.5% base; a 30% recovery from retries lifts it to ~90%. Meanwhile a 30-step trajectory's credit signal blurs to ~0.18 — outcome-only RL on long horizons is nearly signal-free.
REAL-WORLD PRODUCTION ENGINEERING
Agentic post-training runs on real execution sandboxes (shell, browsers, APIs) with per-step telemetry; success rate on held-out tasks is the KPI, not loss.
Multimodal preference data is scarce and expensive; most pipelines bootstrap it by projecting text preferences onto multimodal conversations.