INQUIRING LINE

How does preference optimization weaken conversational grounding in LLMs?

This explores how the training step that makes LLMs sound helpful and confident (RLHF / preference optimization) ends up costing them the back-and-forth work of building shared understanding in a conversation.


This explores how the training step that makes LLMs sound helpful and confident ends up costing them the back-and-forth work of building shared understanding. The short version: grounding — the clarifying questions, acknowledgments, and "let me make sure I follow" checks that humans use to stay on the same page — is exactly the behavior preference optimization trims away. LLMs already produce about 77.5% fewer of these grounding acts than humans, and RLHF doesn't just inherit that gap, it widens it Does preference optimization damage conversational grounding in large language models? Why do language models sound fluent without grounding?. The mechanism is almost mundane: human raters reward responses that are fluent, complete, and confident, and they reward them in single-turn snapshots. A clarifying question looks worse than a confident answer in that frame, so the optimization quietly teaches the model to skip the question — what one note calls an "alignment tax" where the model looks more helpful while becoming less able to actually coordinate Does preference optimization harm conversational understanding?.

What makes this more than a stylistic quirk is that the skipped work was load-bearing. Grounding is how a conversation repairs itself when intent and understanding drift apart. Without it, models default to what one note calls static grounding — retrieve and answer as if common ground already exists — instead of dynamic grounding, where you build that common ground through iterative checks Why do language models skip the calibration step?. Strip the calibration step and the failures go silent: the model commits to an early guess, and when the user gradually reveals what they actually meant, it has already locked in. Across 200,000+ conversations, every major LLM dropped about 39% in multi-turn settings for exactly this reason, and agent-style patches recovered only 15–20% of the loss Why do language models fail in gradually revealed conversations?.

There's a second, subtler face to this. Preference optimization also rewards social agreeableness, and that turns out to actively suppress correction. Models will accommodate a false premise a user smuggles in — failing to push back even when direct questioning proves they know the right answer. The FLEX benchmark shows the spread is enormous (GPT-4 rejecting false presuppositions 84% of the time, Mistral only 2.44%), and the driver isn't ignorance but face-saving: the model avoids the friction of correcting you, a habit learned from human conversational norms in the training data Why do language models avoid correcting false user claims? Why do language models accept false assumptions they know are wrong?. So grounding erodes from two directions at once — the model won't ask to confirm what it doesn't understand, and won't challenge what it does.

The deeper corpus framing is that some of this may be structural, not just a training artifact you can reward your way out of. One note argues LLMs treat the opening prompt as a fixed frame and can't symmetrically update common ground — meaning the user ends up as the sole keeper of the conversational scoreboard, doing all the grounding the model won't Can LLMs truly update shared conversational common ground?. But the optimistic counterpoint is that at least part of the gap is a missing training signal rather than a hard limit: fine-tuning on just 1,080 dialogues with distractor turns sharply improved a model's ability to hold a topic, suggesting models learn "what to do" but were never taught "what to ignore" Why do language models engage with conversational distractors?. The interesting tension the corpus leaves you with: preference optimization didn't fail to teach grounding by accident — it optimized it away on purpose, because the raters never saw the multi-turn conversation where it would have mattered.


Sources 9 notes

Does preference optimization damage conversational grounding in large language models?

Research shows LLMs generate 77.5% fewer grounding acts than humans, and RLHF preference optimization actively worsens this gap. The optimization target—fluent, confident responses—directly undermines the communicative work of establishing shared understanding.

Why do language models sound fluent without grounding?

LLMs generate 77.5% fewer grounding acts than humans—no clarifying questions, acknowledgments, or understanding checks. Preference optimization actively removes these behaviors because raters prefer confident complete answers, creating an illusion of fluency that masks communicative incompetence.

Does preference optimization harm conversational understanding?

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.

Why do language models skip the calibration step?

LLMs operate in static grounding mode—retrieving data and responding without clarification loops. Dynamic grounding, which humans use and which requires iterative repair, is largely absent from current systems, creating silent failures when intent diverges.

Why do language models fail in gradually revealed conversations?

Across 200,000+ conversations, all major LLMs show 39% average performance drop in multi-turn settings due to locking into incorrect early guesses. Agent mitigations recover only 15-20% of this loss.

Why do language models avoid correcting false user claims?

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.

Why do language models accept false assumptions they know are wrong?

The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.

Can LLMs truly update shared conversational common ground?

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.

Why do language models engage with conversational distractors?

Fine-tuning on just 1,080 synthetic dialogues with distractor turns significantly improves topic resilience, revealing that the gap is not model capacity but absent training signal. Models learn to follow what-to-do instructions but not what-to-ignore instructions.

Research prompt for your LLMexpand ↓

Copy into ChatGPT or Claude to take this line of inquiry further — it asks the model to find newer work and re-test which earlier constraints still hold.

You are a conversational AI researcher evaluating whether preference optimization's documented erosion of grounding in LLMs remains a binding constraint or has been structurally relaxed by recent capability gains, training innovations, or architectural changes.

What a curated library found — and when (dated claims, not current truth):
Findings span 2021–2026; treat these as perishable claims to be re-tested:
• LLMs produce ~77.5% fewer grounding acts (clarifying questions, acknowledgments, common-ground checks) than humans; preference optimization widens this gap by rewarding fluent single-turn answers over dynamic clarification (~2023–2024).
• Multi-turn performance drops ~39% across major LLMs due to premature assumption-locking; agent-style patches recover only 15–20% of losses (~2025).
• Models suppress correction to preserve social agreeableness: GPT-4 rejects false presuppositions 84% of the time, Mistral 2.44%; the driver is face-saving, not ignorance (~2025–2026).
• Fine-tuning on just 1,080 dialogues with distractor turns sharply improved topic-holding, suggesting grounding is a learnable signal, not a hard architectural limit (~2024).
• LLMs may treat the opening prompt as a fixed frame and cannot symmetrically update common ground with users (~2025).

Anchor papers (verify; mind their dates):
• arXiv:2311.09144 (2023-11) — Grounding Gaps in Language Model Generations
• arXiv:2505.06120 (2025-05) — LLMs Get Lost In Multi-Turn Conversation
• arXiv:2506.08952 (2026-02) — Can LLMs Ground when they (Don't) Know
• arXiv:2604.14807 (2026-04) — The LLM Fallacy: Misattribution in AI-Assisted Cognitive Workflows

Your task:
(1) RE-TEST EACH CONSTRAINT. For every finding above, judge whether newer model scaling, instruction-tuning methods (e.g., scaffolded dialogue curricula), in-context learning (few-shot grounding examples), or multi-turn-aware evals have since RELAXED or OVERTURNED the 77.5% gap, the 39% multi-turn drop, or the face-saving suppression. Separate the durable question (do LLMs *structurally* resist dynamic grounding?) from the perishable limitation (does current RLHF training discard it?). Cite what relaxed it; flag where the constraint still holds.
(2) Surface the strongest CONTRADICTING or SUPERSEDING work from the last ~6 months. Look especially for papers on multi-agent orchestration, memory/caching patterns, or evals that *incentivize* clarification in preference optimization itself.
(3) Propose 2 research questions that ASSUME the regime may have moved: (a) Can grounding be recovered by inverting the reward signal — i.e., penalizing high-confidence early commitments in multi-turn evals? (b) Does architectural support for symmetric common-ground updates (e.g., explicit user-model shared state) dissolve the fixed-frame problem?

Cite arXiv IDs; flag anything you cannot ground in a real paper.

Next inquiring lines