arXiv Science⌕ Search

arXiv · 2611.19840

Grounding Language Models in Speech Does Not Fix "AI Flavor": Calibrated LLM Judges, Goodhart Traps, and a Preference-Tuning Fix

Abstract

Chinese outputs of large language models carry a perceptible "AI flavor": formulaic sentence frames, stacked written-register connectives, repetition loops, and layout leakage. A popular hypothesis blames text-only training for the missing inner-speech layer of human writing; injecting speech or prosody signals should therefore reduce it. We test it with controlled interventions on a 0.6B model; all three fail: (i) phoneme-interleaved continued pretraining yields no spokenness gain under text-matched controls while raising repetition (+0.13 [+0.03, +0.23]); (ii) prosody-plan conditional generation reaches an adherence of only 0.257 against a pre-registered 0.50 gate; (iii) a rule-based reward built on TTS prosody-feature distance is imperceptible in double-blind evaluation by both humans (8/16) and a calibrated LLM judge (9/16). A root cause is evaluation: single-rater blind evaluation at n~20 only detects extreme (>=85%) preferences. We therefore build a calibrated LLM-judge protocol: the judge passes catch trials on known-answer pairs (16/16; single-text 22/24, same-topic pairs 24/24, versus 0.75 for the human rater), then annotates with two independent judges under A/B swapping, counting disagreement as a tie. Using the judge's agree labels directly as DPO pairs (no reward model), we obtain the first positive result: held-out blind win rate 0.833 [0.699, 0.972], repetition 0.114 to 0.000, unchanged MMLU. We further identify a Goodhart channel: the judge's length preference transduces into output-length collapse via pairwise DPO (median 75 chars vs a 300-char prompt); an SFT-anchor repairs it (w is a style knob: w=0.1 preserves gains, w=0.3 does not). Scaling preference data from 89 to 240 pairs yields no additional judge-visible gain. We release RA-Eval v0.1: a three-layer Chinese AI-flavor evaluation framework with 382 judge-annotated preference pairs.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Dongwu Qian. 2026-11-15. Grounding Language Models in Speech Does Not Fix "AI Flavor": Calibrated LLM Judges, Goodhart Traps, and a Preference-Tuning Fix. https://doi.org/10.5281/zenodo.24237533. https://arxiv.science/abs/2611.19840

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Response Attack: Exploiting Contextual Priming to Jailbreak Large Language Models

Contextual priming, where earlier stimuli covertly bias later judgments, offers an unexplored attack surface for large language models (LLMs). We uncover a contextual priming vulnerability in which the previous response in the dialogue can steer its subsequent behavior toward policy-violating content. While existing jailbreak attacks largely rely on single-turn or multi-turn prompt manipulations, or inject static in-context examples, these methods suffer from limited effectiveness, inefficiency, or semantic drift. We introduce Response Attack (RA), a novel framework that strategically leverages intermediate, mildly harmful responses as contextual primers within a dialogue. By reformulating harmful queries and injecting these intermediate responses before issuing a targeted trigger prompt, RA exploits a previously overlooked vulnerability in LLMs. Extensive experiments across eight state-of-the-art LLMs show that RA consistently achieves significantly higher attack success rates than nine leading jailbreak baselines. Our results demonstrate that the success of RA is directly attributable to the strategic use of intermediate responses, which induce models to generate more explicit and relevant harmful content while maintaining stealth, efficiency, and fidelity to the original query. The code and data are available at https://github.com/Dtc7w3PQ/Response-Attack.

cs.CL↗

Stress-Testing Emotional Support Models: Moving from Homogeneous to Diverse Help Seekers

As emotional support chatbots have recently gained significant traction across both research and industry, a common evaluation strategy has emerged: use help-seeker simulators to interact with supporter chatbots. However, current simulators suffer from two critical limitations: (1) they fail to capture the behavioral diversity of real-world seekers, often portraying them as overly cooperative, and (2) they lack the controllability required to simulate specific seeker profiles. To address these challenges, we present a controllable seeker simulator driven by nine psychological and linguistic features that underpin seeker behavior. Using authentic Reddit conversations, we train our model via a Mixture-of-Experts (MoE) architecture, which effectively differentiates diverse seeker behaviors into specialized parameter subspaces, thereby enhancing fine-grained controllability. Our simulator achieves superior profile adherence and behavioral diversity compared to existing approaches. Furthermore, evaluating 7 prominent supporter models with our system uncovers previously obscured performance degradations. These findings underscore the utility of our framework in providing a more faithful and stress-tested evaluation for emotional support chatbots.

cs.CL↗

Thinking Is Not Telling: Information Disclosure in User-Service LLM Agents

User-engaged LLM agents increasingly operate in service scenarios where task success depends on coordination between the agent, the user, and a stateful environment. In such interactions, the agent often has access to rich information that the user does not observe, such as task policies, tool results, and environment states. This makes agent-user communication a central component of task completion. In this work, we study a communication-level failure mode of user-service agents: internal reasoning is counterproductive for the interactive information disclosure. We introduce information disclosure as a measurable communication mechanism, covering state corrections, available options, constraints, consequences, and proactively retrieved environment information. Motivated by recent advances in LLM reasoning, we evaluate and analyse information disclosure under a controlled setting of agents equipped with a "think" tool, which allows agents to reason internally before responding or acting. Across five models and three user-service scenarios, we first show that enforced thinking has unstable utility. We then analyse response scaling under different think efforts and find that more thinking does not reliably translate into richer information disclosure. Using turn-level counterfactual intervention, we show that injecting golden decision-relevant information significantly intervenes agents to correct action trajectories, indicating that missing disclosure can causally affect the agent's immediate downstream behaviour. We further find that failed trajectories most often lack state corrections and constraints, and that a lightweight disclosure-aware prompt effectively mitigates degradation. These results suggest that the key challenge is not simply making agents think more, but ensuring that reasoning and information are externalised as timely information disclosure.

cs.CL↗