What is the actual reusable unit of reasoning data?
Does post-training reasoning transfer as prompt-response pairs, or as something more complex? Understanding what artifact actually drives gains matters for reproducibility and attribution.
The most useful move in this survey of 150+ post-training studies is a reframing of what reasoning data actually is. The field talks as if the asset being released is a dataset of prompt-response pairs. The primer argues the real reusable unit is a "verifier-bearing feedback interface" whose value is inseparable from six entangled factors: the verifier, the base model, the data lineage, the optimizer, the scaffold, and the inference budget. Change any one and the same "data" produces different gains. The central unresolved question therefore becomes attribution: when a model improves, which part of that interface changed?
This is the connective tissue under several findings the vault already holds separately. When does RL actually extend reasoning beyond pretraining? is exactly the base-model-and-lineage dependency the primer names — gains attributed to "data" are really data-times-headroom. Does RL teach reasoning or just when to use it? is the optimizer-and-scaffold dependency: the interface re-weights existing capability rather than installing new data content. And How do quality, diversity, and complexity affect synthetic data differently? is the construction half of the same problem — a dataset's effect cannot be read off its quality alone because the verifier and budget co-determine it.
The strongest counterargument is that "it's all entangled" can become an excuse for never isolating anything — a survey-level shrug. The primer's defense is that attribution is tractable if releases ship the interface, not just the pairs: report the verifier, the base, the optimizer, the budget, so gains become inspectable, comparable, and testable. For writing, the sharp claim is that the post-training literature's reproducibility crisis is a units problem — people are sharing the wrong object, and benchmark numbers without the interface are uninterpretable.
Inquiring lines that read this note 22
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
How do reward signal structures affect whether models learn true reasoning? What makes decomposed prompts more reliable than monolithic instructions? How do AI systems ground research citations against hallucination? Can recurrence enable transformers to solve problems beyond fixed-depth limitations? Why does fine-tuning for specific tasks degrade broader reasoning capability?- Why do reasoning gains resist clear attribution to specific training changes?
- Can training-time debate work on tasks beyond mathematics and verifiable answers?
- Why does reconstructing hidden thought processes during training improve cross-domain transfer?
- Do identical task structures mean repeated instances or new synthetic samples with same design?
- Can training data organization by capability outperform source or task-based mixing?
- Does the 23-point improvement replicate across different sets of training tasks?
Related concepts in this collection 5
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
When does RL actually extend reasoning beyond pretraining?
Does reinforcement learning genuinely expand a model's reasoning capabilities, or does it merely improve sampling from existing knowledge? This question hinges on whether pretraining provides sufficient foundation and whether RL targets tasks within reach.
grounds (the base-model and lineage dependency the primer names)
-
Does RL teach reasoning or just when to use it?
Does reinforcement learning in thinking models actually create new reasoning abilities, or does it simply teach existing capabilities when to activate? This matters for understanding where reasoning truly emerges.
grounds (the optimizer/scaffold dependency: re-weighting not new content)
-
How do quality, diversity, and complexity affect synthetic data differently?
When training models on synthetic data, do quality, diversity, and complexity each play distinct roles in how well models generalize? Understanding their separate effects could explain why current optimization strategies fail.
extends (construction-side instance of the attribution problem)
-
Do base models find more solutions than post-trained ones?
Explores whether post-training sacrifices solution diversity for single-shot accuracy in agentic tasks, and whether base models with simple prompting can recover that lost coverage at scale.
Evidence for: base models outscoring post-trained ones on coverage complicates attributing reasoning gains to post-training, as A argues
-
Does pretraining loss predict performance after reinforcement learning?
Can we predict how well a language model will perform after RL training based only on its pretraining loss? Understanding this connection could guide how to allocate compute between pretraining and RL stages.
Evidence for: pretraining loss/tokens predicting post-RL performance shows base-model lineage, one of A's attribution factors, actually drives RL gains
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- A Primer in Post-Training Reasoning Data: What We Know About How It Works
- Escaping Model Collapse via Synthetic Data Verification: Near-term Improvements and Long-term Convergence
- An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models
- On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models
- Local Coherence or Global Validity? Investigating RLVR Traces in Math Domains
- Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
- OpenThoughts: Data Recipes for Reasoning Models
- Eliciting Reasoning in Language Models with Cognitive Tools
Original note title
the reusable unit of post-training reasoning is not a prompt-response pair but a verifier-bearing feedback interface — which is why reasoning gains resist attribution