Model Post-Training Bottleneck

Definition

Model post-training bottleneck is the constraint that useful model behavior after pretraining depends on aligned data, evaluation, expert judgment, and iteration, not merely on having a pretrained base model or running a fine-tuning job.

Current Synthesis

Across the bounded sources, post-training is the layer where latent model capability becomes reliable behavior for a task. 张奇 frames it as matching the base model’s existing knowledge and behavior limits; 孟繁青 frames it as a data, environment, and verification problem inside model labs; 逯雨鑫 / 逯雨昕 makes the same bottleneck visible at individual scale, where training can be cheap but data audit and failure diagnosis consume most of the work.

The current judgment is that post-training bottlenecks scale down as well as up. Frontier labs may have expensive tacit recipes, expert labels, RL systems, and internal services, while individuals can use SFT and QLoRA for narrow improvements. In both cases, the scarce resource is not the button that starts training. It is knowing what behavior should change, whether the data really teaches that behavior, and whether evaluation catches regressions.

Key Claims

• Post-training must match the base model’s actual knowledge, capacity, and behavior distribution.
• Data that looks high quality can fail if it pulls the model away from the intended target or exceeds the student model’s capacity.
• The bottleneck often sits in data construction, manual audit, task design, and evaluation rather than in training runtime.
• Agentic work makes the bottleneck harder because traces must include tools, environment feedback, failure recovery, and verification.
• Narrow, low-cost post-training can produce real domain gains, but it does not imply general frontier-model progress.
• Serving, latency, concurrency, and model-routing costs remain separate constraints after a post-trained model improves.

Evidence

Data matching and latent capability

• Vol.114 AI的2025和DeepSeek们的未来 | 对谈复旦张奇教授 argues that post-training works best when it unlocks knowledge already present in pretraining, while mismatched supervised data can fail or disturb other behavior.

Environment and verification

• 从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青 treats post-training leverage as task construction, leakage avoidance, difficulty verification, environment-based benchmarks, and model-improvement checks.

Individual-scale bottleneck

• 一个人、两周、数百美元，如何训出登顶 Hugging Face 的模型 | 对谈研究员逯雨鑫 reports a few-hundred-dollar, days-to-weeks project where most time went to data review and benchmark-error iteration rather than the hours-long training runs.

Agent and data-market extension

• 从蒸馏到合成数据到 RSI，模型竞争的下一个焦点是什么？｜对谈 Evolvent AI 联创孟繁青 links the bottleneck to Synthetic Agent Data, Environment-Based Agent Benchmarks, and RSI Data, while 一个人、两周、数百美元，如何训出登顶 Hugging Face 的模型 | 对谈研究员逯雨鑫 shows real trajectories and manual audit as the small-team analogue.

Counterevidence & Qualifications

• The newer source qualifies the idea that post-training is always expensive: narrow improvements can be cheap when the target is specific and the builder accepts regression risk elsewhere.
• The sources do not prove a universal data-size rule; useful data volume depends on model size, target behavior, evaluation quality, and how much relevant ability is already latent.
• Strong benchmark movement can still be overfit or too narrow unless checked against real workflow value and regressions.
• Post-training is not a substitute for frontier pretraining when the base model lacks the underlying knowledge or capability.

What Changed

• Migrated the page to synthesis-v1 and reorganized evidence by claim rather than source arrival.
• Added the individual-builder case showing that the same bottleneck appears even when training is cheap.
• Clarified the boundary between narrow, low-cost improvement and general frontier capability.

Related Concepts

• Data-First Post-Training / 数据优先后训 - operational discipline for addressing this bottleneck.
• Low-Cost Model Post-Training / 低成本模型后训 - small-team route that makes the bottleneck visible at low budget.
• Supervised Fine-Tuning / SFT - common method that still depends on data and evaluation quality.
• Model Distillation / 模型蒸馏 - teacher-student transfer pattern constrained by the same data and capacity limits.
• Agent Post-Training - harder agent-specific form involving tools, traces, and recovery behavior.
• AI Verification - validation layer needed to detect real improvement and regressions.