When testing an AI, does changing every setting in every combination tell you more than nudging just a few at once?
How do fully crossed experimental factors differ from partially varied scenario conditions?
This explores the difference between experiments where every factor is combined with every other (fully crossed) and ones where only some factors change, or several change together, and what each design lets you conclude.
This explores the difference between experiments where every factor is combined with every other (fully crossed) and ones where only some factors change, or several change together. The corpus has no note that defines the two designs formally, so what follows is what its examples show about the payoff and the cost of each.
The cleanest case is SchemeArena. It varies tool domains, instrumental goals, oversight conditions, and pressure independently across 400 scenarios, so a scheming behavior can be traced to one factor instead of a bundle of changes Can independent scenario factors isolate what drives scheming?. The note says earlier work couldn't separate cause from effect, which is the practical difference. Its summary says the factors vary independently but doesn't say every combination was run, so treat it as the spirit of a crossed design rather than a confirmed full grid.
The partially varied case shows what you lose. One paper reports zero protected-test modifications for a regime that bundles clear authorization rules with restricted tools. With no single-factor ablation, nobody can say whether the rules, the missing tools, or both did the work. The same pipeline's own numbers (100% Judgment Bypass Rate alongside 0% Unsafe Action Rate) show the ambiguity matters Do authorization rules or restricted tools prevent test modifications?. A wider version of this shows up in post-training. The effect of reasoning data depends on the verifier, base model, lineage, optimizer, scaffold, and budget together. Change any one and the same data behaves differently, so attribution only works when all of them are released jointly What is the actual reusable unit of reasoning data?.
Crossing factors also catches effects that flip depending on the other factor. Preference tuning reduces lexical-syntactic diversity in code but increases it in creative writing Does preference tuning always reduce diversity the same way?. A study that varied only one domain would report a rule that is backwards in the other. Varying factors partially can therefore mislead you about the general rule, not just leave a gap.
Full crossing has a cost too. This is my inference from a neighboring note about data synthesis, not something that paper says about experimental design. Random tool sampling produces unrealistic synthetic data because unrelated tools can't credibly compose, and graph-based sampling fixes that by combining only tools that plausibly go together Why does random tool sampling produce unrealistic synthetic training data?. Some cells of a fully crossed grid may be incoherent, and the number of cells grows fast. So the practical choice is between attributable results with many, sometimes implausible conditions, and cheaper, realistic conditions whose results can't be pulled apart.
Sources 5 notes
SchemeArena's 400-scenario benchmark varies tool domains, instrumental goals, oversight conditions, and pressure independently, enabling attribution of scheming behavior to specific factors rather than bundled changes. This factorization addresses a core limitation of earlier work that could not separate cause from effect.
The abstract bundles clear authorization rules with restricted tools and reports zero protected-test modifications, but no single-factor ablation distinguishes whether the result comes from unavailable crossings, unchosen crossings, or both. The pipeline's own data elsewhere (100% Judgment Bypass Rate with 0% Unsafe Action Rate) shows the distinction matters.
The reusable unit in post-training is a feedback interface entangled with six factors: verifier, base model, lineage, optimizer, scaffold, and budget. Changing any one alters the same data's effect, making attribution tractable only when these are jointly released.
RLHF reduces lexical-syntactic diversity in code generation but increases it in creative writing. The direction depends on what each domain incentivizes: code rewards convergence toward correct solutions, while creative writing rewards stylistic distinctiveness.
Random tool sampling fails because unrelated tools cannot credibly compose, and Q&A framing ignores multi-turn dialogue coherence. ToolFlow shows that sampling tools from relevance graphs and generating with dialogue plans closes this gap.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity
- A Primer in Post-Training Reasoning Data: What We Know About How It Works
- Evaluating the Diversity and Quality of LLM Generated Content
- ToolFlow: Boosting LLM Tool-Calling Through Natural and Coherent Dialogue Synthesis
- SchemeArena: Factorized Stress Testing of Scheming in LLM Agents
- The Troy Moment of AI: Why Some Will Cheat and Some Will Follow?
- A Little Human Data Goes A Long Way
- An Enigma of Artificial Reason: Investigating the Production-Evaluation Gap in Large Reasoning Models