What makes some frictions negligible while others block entire pathways?
This explores why some errors, costs, or interferences get harmlessly absorbed while others compound or sit at chokepoints that derail an entire process — and the corpus answers it less as a question about size than about position and propagation.
This explores why some errors, costs, or interferences get harmlessly absorbed while others bring a whole process to a halt — and the corpus suggests the deciding factor is rarely the size of the friction. It's whether the friction *propagates*, where it *sits*, and whether the system can *wall it off*. A small error in an isolated spot stays small. The same error in a chokepoint, or one that feeds the next step, snowballs.
The clearest case is compounding. A study decomposing chain-of-thought reasoning found that genuine reasoning exists but accumulates error with every step, so what looks negligible per-step becomes fatal over a long chain What three separate factors drive chain-of-thought performance?. Relatedly, a reasoning trace can be locally coherent — each step plausibly follows the last — and still be globally invalid, because friction that's invisible at the seam between two steps quietly breaks the whole proof Does RLVR actually improve mathematical reasoning or just coherence?. This is exactly why measuring friction *locally* matters: step-level confidence catches breakdowns that a global average smooths over, letting you stop a doomed trace early instead of discovering the failure only at the end Does step-level confidence outperform global averaging for trace filtering?.
The second factor is structural position. Some frictions block pathways because they sit where two capabilities are forced to share one channel and pull against each other. GUI agents stall when planning and grounding are bundled into one policy with opposing optimization needs — the fix is an intermediate interface that separates them so each can move without obstructing the other Why do planning and grounding pull against each other in agents?. The same logic appears in fine-tuning: multi-task interference becomes a real blocker only at the parameters that several tasks contend over, and isolating those core regions while freely merging the rest turns a pathway-blocking conflict into a negligible one Can isolating task-specific parameters prevent multi-task fine-tuning interference?.
That points to the third factor — whether friction can be compartmentalized at all. The reason LLMs can't keep contexts cleanly separate is structural: they process everything as one token string with no walls between domains, so a friction in one place bleeds into all the others, and every mitigation just relocates the failure How do LLMs balance remembering context versus keeping it separate?. The same theme shows up in an unexpected place: removing a spurious cue *helps* when the task is to ignore a distractor, but *hurts* when the real job is integrating conflicting signals — so the friction you can safely delete versus the one you can't depends entirely on whether it's a separable nuisance or load-bearing input Why does removing spurious cues sometimes hurt model performance?.
The through-line worth taking away: a friction is negligible when it's isolated, detected locally, and doesn't feed forward; it blocks the whole pathway when it compounds across steps, sits at a contested junction, or can't be walled off from everything downstream. The interesting design move, across all these papers, is the same — turn a blocking friction into a negligible one by changing its *position* (separate the channels, isolate the parameters, measure per-step) rather than by shrinking the friction itself.
Sources 7 notes
A shift cipher study decomposed CoT into three independent factors: output probability alone swings accuracy from 26% to 70%, memorization matches pre-training frequency patterns, and genuine reasoning exists but accumulates error with each step. This resolves the reason-or-memorize debate by showing LLMs do both simultaneously.
RLVR post-training measurably reduces logical errors between adjacent reasoning steps, but locally coherent traces can still be globally invalid proofs. The improvement is structural rather than semantic.
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
AutoGLM's research shows planning and grounding have opposing optimization requirements that pull against each other when bundled in one policy. An intermediate interface that separates them lets each capability be developed and optimized independently while still composing into a complete agent.
Research shows that identifying core parameter regions per task, clustering overlapping tasks, and freezing core parameters while geometrically merging non-core parameters consistently outperforms standard multi-task fine-tuning. Temporal task scheduling alone proves insufficient without explicit structural parameter isolation.
Because LLMs process conversation as a single token string without compartmentalized memory, they cannot maintain separate contexts the way humans do. Existing mitigations like compression, longer windows, and retrieval all introduce new failure modes and cannot replicate human compartmentalization.
Removing spurious cues degrades performance in heuristic override tasks, opposite to shortcut learning predictions. The failure mode is integrating conflicting signals rather than ignoring distractors—a frame problem, not feature selection.