Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.08055

Language Carries the Expert's Impression: Instrument-Anchored LLM Judges Transfer Counseling-Quality Assessment and Beat In-Domain Training

Tobias Hallmen·Elisabeth André
View original source ↗Read source document ↗Export metadata ↓

Abstract

Automatic assessment of communication quality in dyadic counseling conversations is bottlenecked by data: expert-rated corpora are small and expensive to grow. We study cross-domain transfer of expert overall-impression prediction across three German corpora of simulated counseling (two general-practice medical, one school-related parent-teacher; $n=195$ expert-rated sessions, one corpus after scale equating). Training on the other domains beats training in-domain: leave-one-domain-out transfer reaches nested Spearman $ρ= 0.54$ against $\le 0.48$ within the target domain, a paired session-level gap of $+0.15$ that holds at $+0.12$ when the training-set sizes are matched, so it is not simply data volume. The decisive features are session-level construct scores from small open-weight LLMs reading the two-speaker transcript, with the constructs largely derived from the experts' rating instruments: the instrument-derived battery lifts a single judge from $0.32$ to $0.41$ over generic dialogue qualities, judges from three model families ensemble to $0.51$ language-only, and a nonverbal-dyadic block adds $+0.03$ more, not separable from noise at this sample size. We also price the recording setup: one corpus lost its per-speaker audio, 16% of its diarised segments carry the wrong speaker, and repair is worth $+0.07$ there. At practically attainable corpus sizes, the expert's overall impression is carried by what is said, and by other communication programs' data more than by one's own.

Explore related subjects

cs.CLcs.LG

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Tobias Hallmen, Elisabeth André. 2026-10-06. Language Carries the Expert's Impression: Instrument-Anchored LLM Judges Transfer Counseling-Quality Assessment and Beat In-Domain Training. https://arxiv.org/abs/2610.08055

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Document Optimization for Black-Box Retrieval via Reinforcement Learning

Generative large language models (LLMs) are increasingly used as inference-time components in retrieval pipelines, for tasks such as query rewriting and document reranking. However, these online approaches place costly autoregressive computation directly on the latency-critical retrieval path. We explore an alternative axis: using LLMs to improve documents instead, rewriting them into better representations and shifting computation offline. Yet producing a useful document rewrite is not straightforward: retrieval is inherently discriminative, so an effective rewrite must make a document more similar to relevant queries than competing candidates under the retriever's notion of similarity. We therefore formulate document transformation as an optimization problem, directly training an LLM or VLM to produce rewrites that improve retrieval. Our approach, DocOpt, uses GRPO with retriever ranking improvements as rewards, requires only black-box access to retrieval ranks, and applies across single-vector, multi-vector, and lexical retrievers. We evaluate zero-shot LLM rewriting and DocOpt on code and visual retrieval tasks, finding that document rewriting can improve retrieval and that optimizing rewrites yields further gains. For example, OpenAI text-embedding-3-small achieves 58.35 nDCG@5 on average with direct retrieval; zero-shot rewriting improves this to 60.83 with GPT-5.4-mini, 63.75 with Claude Haiku 4.5, and 64.23 with Qwen3. DocOpt further improves performance to 67.94, surpassing the 6.5X more expensive text-embedding-3-large retriever at 66.15.

cs.CL↗
arXiv2026

How Far Do Auto-Interpretation Labels Generalize: A Controlled Study Across Languages, Scripts, and Rewordings

Sparse autoencoder (SAE) features are increasingly used to interpret language models, with auto-generated natural-language labels serving as the primary interface for understanding what each feature represents. We ask whether these labels generalize: does a feature labeled for a concept actually track that concept across languages and scripts? Using Serbian digraphia as a controlled testbed -- the same language written in both Latin and Cyrillic via deterministic transliteration -- we first find that SAE feature sets activated by the same content in different languages, scripts, and wordings share substantial overlap (mean Jaccard 0.39 vs 0.13 random baseline, peaking at 0.57), suggesting genuine cross-lingual semantic features. We then test whether auto-interpretation labels keep pace. They often do not: features whose labels describe semantic content miss the same meaning in Serbian up to 4$\times$ more often than within English, and miss Serbian Cyrillic more than Serbian Latin -- two scripts that are deterministic transliterations of each other. The gap grows with network depth, yet the labels give no indication that they fail. These results suggest that auto-interpretation labels reflect a feature's behavior on the languages and scripts a model has seen most in training, rather than the concept itself.

cs.CL↗
arXiv2026

Progressive Disclosure for LLM-Maintained Wiki Knowledge Bases: a Preregistered Ablation

LLM agents now often answer questions from knowledge bases they help maintain. A common intuition says progressive disclosure should make this cheaper. Instead of loading one large index, the agent reads a compact catalog and one-line page summaries, then opens only the pages it needs. We tested that intuition in a preregistered study on a real 709-page markdown knowledge base maintained by an LLM. We retrofitted it for progressive disclosure and built four versions that differ only in how the agent reaches the pages. The pages themselves are identical in every version, so any difference comes from the access structure alone. Each version was tested three ways, with the agent following a set protocol, choosing its own path, or made to load the catalog first. A judge from a different model family graded the answers blind against verified reference answers. A preparatory pilot changed the question. A capable agent never loaded the large index at all. It worked out from the question where a page was and read it directly. The saving we set out to measure did not exist for such an agent, so we made answer quality the primary outcome. Quality held. Answers from the retrofitted knowledge base were as good as answers from the original, within a margin we set in advance. Two limits apply. Our human rater and the model judge agreed far less than the plan required, so the quality result rests on the judge, backed by sensitivity checks. Quality was also not shown to hold when the agent was forced to load the catalog first, or on the two most reliably graded criteria under a stricter test. Cost fell clearly in every condition we tested, and the retrofitted version cited fewer pages and took fewer tool turns per answer.

cs.CL↗