When an AI method adds two things at once, which one actually earns the improvement, or does it need both?

This explores an ablation-style question: when a method feeds concepts back to the model and also adds an auxiliary training objective, which part produces the gain? None of the retrieved notes runs that exact comparison, so the corpus can't say which side wins for the method you have in mind. What it does have is several studies that split "the content going in" from "the structure or training around it", and they point the same way. The content that looks like it is doing the work often isn't.

Two notes show the pattern. Chain-of-thought prompts with logically invalid reasoning performed nearly as well as valid ones, so the gain came from the form of reasoning, not the correctness of its content (Does logical validity actually drive chain-of-thought gains?). Skills are similar. Across 8,135 trials, 65.7% of skill benefits came from procedural anchoring and only 4.5% from injecting missing knowledge (Do skills teach procedures or inject missing facts?). If fed-back concepts behave like these, the concepts may matter less than the fact that the model is trained or prompted into a structured loop. A third note has the same flavor. Extended thinking hurt vanilla models through self-doubt but helped after RL training, with the thinking content unchanged (Does extended thinking help or hurt model reasoning?). The objective decided whether the content was useful.

The case for the fed-back concepts rests on what feedback carries. Feedback has an evaluative part (how well it went) and a directive part (what to change), and a scalar-style objective captures only the first. The directive part needs a channel that carries specifics (Can scalar rewards capture all the information in agent feedback?). If your fed-back concepts are that directive channel, an auxiliary objective alone can't replace them. Where the concepts come from also matters. Self-improvement stalls without external anchors such as judges, tools, or user corrections (Can models reliably improve themselves without external feedback?). Concepts the model generated itself should help less than concepts from outside it. Repeated cycles of trying, getting feedback, and incorporating it were also the strongest predictor of success on very long tasks (What predicts success in ultra-long-horizon agent tasks?).

The case for the objective alone is that a training loss reshapes what the model learns, whatever inputs it sees. Utility-weighted loss improved decisions but weakened representation learning, and a symmetric loss with post-hoc adjustment beat it on the same target (Can utility-weighted training loss actually harm model performance?). An auxiliary objective could likewise produce gains through gradient signal alone, with the fed-back concepts as passengers.

A clean answer needs a four-way test: neither component, concepts only, objective only, and both. The objective-only arm should still get concept-shaped inputs, but shuffled or random ones, so the concepts' content is separated from the mere presence of extra text. Even then, the combined gain can fall short of what the parts promise. One study found people captured only about half of an LLM's item-level accuracy gain (Why does assisted accuracy capture only half the LLM gain?). Check the result on held-out tasks too, as the AIDE2 work did, so the gain isn't specific to the selection set (Do AIDE2's improvements transfer to unseen tasks?).