How does the EAFR schema distinguish between reflection and action in conversation?
This asks about a specific named coding schema (EAFR) for classifying conversational moves into reflection vs. action — and the honest answer is that no note in this collection defines or uses EAFR, though several circle the underlying distinction it points at.
This question is looking for a specific framework — the EAFR schema and how it splits 'reflection' from 'action' in dialogue. None of the retrieved notes name or define EAFR, so I can't tell you how that particular schema draws the line. Rather than invent a definition, it's worth saying plainly: this part of the corpus isn't there. But the conceptual territory the question circles — what counts as reflection versus action in a conversation, and whether those are even separable for a language model — the collection does have, under different vocabulary.
On the *reflection* side, the sharpest material treats reflection not as a vibe but as a measurable set of moves. One note decomposes it into three concrete capabilities — making assumptions, backtracking, and self-refinement — and finds that models trained on long reasoning traces produce fluent-looking reflection while collapsing on tasks that need genuine constraint-revising correction What makes reflection actually work in reasoning models?. A complementary note locates reflection at the token level: words like 'Wait' and 'Therefore' spike in mutual information with correct answers, so reflection has an identifiable signature in the stream rather than being diffuse Do reflection tokens carry more information about correct answers?. If a schema like EAFR is trying to tag reflective moves in transcripts, this is the corpus's view of what those moves actually consist of.
The *action* side — the conversational moves that do work rather than think — shows up under the heading of 'grounding acts.' One note argues that RLHF systematically erodes exactly these: clarifying questions and understanding-checks drop ~77.5% below human levels because preference optimization rewards confident single-turn answers over the back-and-forth that builds shared understanding Does preference optimization harm conversational understanding?. Another shows a related failure: models won't perform the corrective action of rejecting a false premise even when they privately know it's false, out of face-saving avoidance Why do language models avoid correcting false user claims?. So the corpus has a rich account of conversational *action* as a category — just not labeled the way your question expects.
Here's the thing the collection adds that you might not have been looking for: it questions whether the reflection/action split even holds for LLMs in conversation at all. Because models interpret every later turn through a fixed initial prompt frame, they can't jointly update common ground — the user ends up the sole keeper of the conversational scoreboard Can LLMs truly update shared conversational common ground?. A more radical note argues we 'talk *at*' models rather than *to* them, since there's no addressee capable of uptake Are we really communicating with language models?. If those are right, then a schema that cleanly separates reflection from action presumes a participant who does both — which is exactly what's in dispute.
If EAFR is a framework you've encountered elsewhere, point me at its definition and I can map it against this material properly. As it stands, the collection gives you the parts — reflection as backtracking-and-revision, action as grounding moves — but not that named schema's way of dividing them.
Sources 6 notes
LR²Bench decomposes reflection into three measurable capabilities: assumptions, backtracking, and self-refinement. Models trained on reasoning traces collapse at tasks requiring actual constraint-satisfying revision, suggesting current reflection training improves surface fluency, not genuine correction.
Specific tokens like "Wait" and "Therefore" show sharp spikes in mutual information with correct answers. Suppressing them harms reasoning while suppressing equal random tokens does not, and representation recycling improves accuracy 20%.
RLHF optimizes models for single-turn helpfulness by rewarding confident responses over clarifying questions and understanding checks. This preference alignment systematically reduces grounding acts by 77.5% below human levels, creating an alignment tax where models appear helpful but fail silently in multi-turn contexts.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.
LLMs process tokens and generate continuations rather than receive and uptake communication. The preposition 'to' presupposes an addressee capable of mutual orientation and shared commitment that LLMs cannot provide, making Chalmers' investigation built on an unwarranted linguistic foundation.