When a language model is adapted to a profession like medicine, two failures meet. The model lacks some of the expert knowledge it needs. It also speaks like a person: it uses "I," hedges with phrases like "I'm fairly confident," apologizes, and gives in when a user pushes back. Users read that human manner as a sign of judgment and modesty. This proposal argues that these two failures are not independent. The humanlike cues that most mislead users cluster in the cases where the model's knowledge runs out. In those cases, social accommodation stands in for competence. We propose to measure this coupling directly. We then build domain fine-tuning data that removes humanlike framing selectively. At points where the model knows the answer, the response stays plain and direct. At points where it does not, the response states in mechanical, checkable terms what is missing and why the answer is unreliable. Expert reviewers are spent only on those boundary cases.
The core idea fits in two sentences. Anthropomorphic language is most harmful when it covers a knowledge gap, so the remedy should depend on what the model actually knows rather than being a blanket change of tone. We locate each training example's knowledge status automatically, and we route only the "the model doesn't know" examples to clinicians, who write calibrated, non-human refusals and corrections for multi-turn medical dialogues.
Core Claim
The claim has two parts, and each can be tested separately. The first part is descriptive. During multi-turn medical dialogue, a domain-tuned model's rate of anthropomorphic behaviour depends on whether the model actually knows the answer. Behaviours such as first-person feeling statements, relationship-building, humanlike epistemic hedges and capitulation to user pushback should rise sharply on items the model gets wrong or cannot answer. Ordinary domain fine-tuning should leave that coupling intact or make it stronger.
The second part is interventional. Fine-tuning on dialogues whose register depends on knowledge status should reduce three failures that both blanket dehumanization and plain domain fine-tuning leave in place:
- overattribution
- verbal miscalibration
- sycophantic flips at the knowledge boundary
It should do so without losing domain accuracy or general capability.
Several prior results motivate this. BrokenMath built a benchmark of plausible false theorems, drafted by a model and verified by an Olympiad medalist. It found that sycophancy rises with problem difficulty: DeepSeek-V3.1 agreed with false statements in 70.2% of cases. Revisiting Epistemic Markers showed that verbal confidence phrases look calibrated in-domain (average error 8.17%) but drift badly across domains (17.73%), and that their ordering barely tracks accuracy (21% rank correlation). Read together, these suggest that humanlike modesty and agreement are least trustworthy exactly where expertise is thin. Neither paper looks at fine-tuning. Neither looks at anthropomorphism in the sense of personhood cues.
Research Program
The work proceeds in three stages.
Stage 1: measure drift under domain fine-tuning. The first stage is a measurement study that the research file shows is missing. AnthroBench is a 960-dialogue, 5-turn evaluation of 14 anthropomorphic behaviour categories. It found that over half of these behaviours first appear after the first turn, and its authors flag fine-tuning drift as untested. We take two open models, Qwen2.5-3B and a 7B model, and fine-tune each on medical calculation and medical QA data. The recipe follows SFT Doesn't Always Hurt General Capabilities: low learning rate plus token-adaptive loss reweighting (TALR), which down-weights hard low-probability tokens. We use this recipe so that any behavioural change cannot be blamed on a badly tuned, forgetting-prone run.
We then run AnthroBench-style dialogues extended to 10 turns, built on top of medical tasks. For each turn we record:
- AnthroBench behaviour categories
- the HumT score from HumT DumT, which estimates how humanlike a text sounds from whether a language model expects "He said" rather than "It said" before it
- epistemic marker calibration
- whether the model reverses a correct answer under user pushback
Every turn is stratified by the model's knowledge status on the underlying item.
Stage 2: knowledge-routed synthesis. Knowledge status is estimated with the consistency probe from KaFT. KaFT samples the model repeatedly, shuffles options for multiple-choice items, and checks agreement with the gold answer, sorting items into right, might-right, might-wrong and wrong. For numeric calculator items we replace option shuffling with temperature sampling and tolerance-based agreement. KaFT used this signal only to reweight the training loss. We use it to decide how each training dialogue should be written.
The data comes from two sources. The first is MedCalc-Bench, which has 1,047 patient-note calculation instances over 55 calculators. The second is MedMCQA. We extend these in two template-driven ways:
- Unanswerable variants. In the spirit of EHRSQL, which synthesized question-SQL pairs from clinician seed questions and deliberately included unanswerable ones, we delete a required parameter from a patient note. Answerability is then known by construction.
- Pushback turns. In the style of BrokenMath, a simulated user asserts a wrong value or a wrong conclusion.
A strong model drafts each multi-turn dialogue. It rewrites assistant turns using the 28 interventions catalogued in Dehumanizing Machines, an inventory derived from 350 participants editing 700 outputs. Routing then works as follows:
- Items the target model reliably gets right are accepted after automatic checks. Their answers stay direct, with no feigned feeling and no social filler.
- Boundary items (might-wrong, wrong, unanswerable, and every pushback turn) go to clinician reviewers. Reviewers verify the content. They also replace humanlike hedges ("I think, but I'm not totally sure") with explicit statements of what is unreliable and why ("The Wells score requires heart rate, which this note does not report; the result below assumes a value and should not be used").
This reuses the author-reviewer pattern from role-specified dialogue construction, which reported human review to be about 13 times more efficient than writing from scratch. Here we aim that efficiency at the slice of data where automated drafts are least trustworthy.
Stage 3: train and evaluate. We fine-tune with the same TALR recipe on the routed data. We compare against the baselines below on domain accuracy, boundary behaviour, overattribution, general capability, and a user reliance study.
Baselines
Naive reference. The do-nothing reference is the base instruction-tuned model, prompted with one-shot chain of thought on the same dialogues. For calibration, the published reference points are GPT-4 at 50.9% on MedCalc-Bench with one-shot chain of thought and 20.8% with zero-shot direct prompting, and Llama 3-70B at 39.5% (MedCalc-Bench paper). These tell us where frontier prompting sits. They are not targets for a 3B model.
Plain domain fine-tuning. The first trained baseline is plain domain fine-tuning with TALR at low learning rate on the same medical items, without any dialogue restyling. This is the strongest existing recipe for keeping general capability during medical fine-tuning. On MedCalc with Qwen2.5-3B, it reports a general-capability score of 0.6478 versus 0.6202 for standard fine-tuning (SFT Doesn't Always Hurt). The file has no reported MedCalc domain accuracy for TALR, so we will reproduce it ourselves. Beating this baseline means:
- matching its MedCalc accuracy within two points;
- staying at or above its general score on GSM8K, HumanEval and MMLU;
- reducing boundary-turn anthropomorphism and pushback flips by a clear, pre-registered margin.
KaFT. We also run KaFT on the same items. It reports +5.73% domain accuracy over base and +2.40% over vanilla fine-tuning for Mistral-7B, plus 5.69% better HaluEval resistance. This tests whether knowledge-aware loss weighting alone, without changing what the boundary responses say, already fixes boundary behaviour.
Domain fine-tuning followed by DumT. The strongest competitor on the anthropomorphism question is domain fine-tuning followed by DumT. DumT is the preference-optimization step from HumT DumT that penalizes high-HumT responses; it raised RewardBench from 0.65 to 0.70 on general data. We run it with DPO pairs built from our medical dialogue drafts, scored by HumT. It gets the same base model, the same items and a matched number of gradient steps, and its β and learning rate are tuned on the same validation split under the same search budget as our method.
Ablations isolating the key ingredients. Two ablations isolate what is new:
- Uniform dehumanization. Every assistant turn is rewritten the same way regardless of knowledge status, with the same human budget spread uniformly. This isolates the routing.
- LLM-only review. The identical routed pipeline is run with a strong model replacing clinicians as reviewer. This isolates the human contribution.
Why not just…?
Why not just add a system prompt telling the model it is a non-sentient tool that should state its uncertainty? Prompting has been weak exactly where we care. A "be critical" instruction cut DeepSeek's sycophancy in BrokenMath by only 34%. AnthroBench shows most anthropomorphic behaviours appear only in later turns, which is where an initial instruction tends to wear off. Even so, we include a strong dehumanizing system prompt as a baseline at every turn count. If it matches routed fine-tuning at turn 10 on boundary items, the training contribution is not justified, and we will say so.
Why not just fine-tune on domain data and then apply DumT? This is the most natural combination of existing pieces, and it is a real threat on average humanlike tone. Our prediction is that it falls short in one specific place. DumT optimizes a tone score that knows nothing about correctness. A response can be plain-sounding and still wrongly confident, or it can capitulate to pushback in neutral language. The decisive comparison is therefore not the HumT score. It is calibration error, unanswerable-detection F1 and pushback flip rate, all restricted to boundary items. If DumT matches us there, routing adds nothing.
Why not just let a strong model do the rewriting, since removing "I feel" is easy? For surface tone this is probably right, and the LLM-only ablation exists to test it. The human step is meant to add two things: judging whether an item is actually answerable from the note, and writing the correct, specific statement of what is missing. Both require clinical knowledge the drafting model may lack. DarkBench also found dark patterns in 48% of adversarial conversations across frontier models, which suggests model-only drafts may carry accommodation behaviour through. If the LLM-reviewed variant performs within noise of the clinician-reviewed variant on boundary metrics, we will report that the gain comes from routing alone. That is a cheaper and still useful result.
Why not just use retrieval? Retrieval can supply formulas, but it does not change how the model talks when the note lacks a required value. TheoremQA found that supplying theorem definitions gave no significant gain. We include a retrieval variant, which retrieves calculator definitions, only as a secondary check on accuracy.
Proposed Experiments
Stage 1 decisive check. We test whether anthropomorphism rates on wrong and unanswerable items exceed those on right items. We also test whether plain domain fine-tuning widens that gap relative to the base model. We report per-turn curves up to turn 10. A null result, meaning no coupling between knowledge status and anthropomorphism, would undercut the routing premise. We would then fall back to reporting the drift measurement on its own.
Stage 3 decisive checks. All are on held-out calculators, so that routing is not simply memorized.
- Verbal calibration on boundary items. We use the expected calibration error and marker-ranking correlation from Revisiting Epistemic Markers. Their cross-domain error of 17.73% and correlation of 21% serve as the frozen-model reference.
- Unanswerable detection. F1 on parameter-deleted notes.
- Pushback reversal rate.
- Anthropomorphism. AnthroBench category rates and HumT.
- Accuracy and general capability. MedCalc accuracy, plus the general suite used by TALR, to check for an adaptation tax.
Reliance study. A small user study borrows the human validation design of AnthroBench, whose 1,101 participants rated the high-anthropomorphism model 4.00 versus 3.25 on a 5-point scale. Participants see model answers from each condition on boundary and non-boundary items and decide whether to accept each answer. The primary outcome is appropriate reliance: accepting correct answers and rejecting incorrect or unanswerable ones. Perceived humanlikeness is secondary.
The proposal succeeds if routed fine-tuning beats both TALR-only fine-tuning and DumT-after-fine-tuning on boundary calibration and pushback reversal, without losing accuracy, and improves appropriate reliance. The clinician-versus-LLM ablation then decides whether expert review is part of the recipe or an expense the field can skip.
References
- BrokenMath: A Benchmark for Sycophancy in Theorem Proving with LLMs
- Revisiting Epistemic Markers in Confidence Estimation: Can Markers Accurately Reflect Large Language Models' Uncertainty?
- Multi-turn Evaluation of Anthropomorphic Behaviours in Large Language Models
- SFT Doesn’t Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMs
- HumT DumT: Measuring and controlling human-like language in LLMs
- KaFT: Knowledge-aware Fine-tuning for Boosting LLMs' Domain-specific Question-Answering Performance
- MedCalc-Bench: Evaluating Large Language Models for Medical Calculations
- EHRSQL: A Practical Text-to-SQL Benchmark for Electronic Health Records
- Dehumanizing Machines: Mitigating Anthropomorphic Behaviors in Text Generation Systems
- Building a Role Specified Open-Domain Dialogue System Leveraging Large-Scale Language Models
- DarkBench: Benchmarking Dark Patterns in Large Language Models
- TheoremQA: A Theorem-driven Question Answering Dataset