PaperlayerView exact revision on arXiv

Source-linked AI summary

Clinical Concept Centers in LLMs

Aishik Nagar, Abhishek Vaidyanathan, Arun-Kumar Kaliya-Perumal, Elijah Tzen Hsuen Boey, Stefan Winkler

arXiv:2610.02829v2cs.CLcs.LG

TL;DR

Whether clinical concept centers exist in LLMs and how to locate them is an open question, despite the stated need for mechanistic evaluation of clinical models. This study identifies centers in open-weight models and finds they support evaluation and clinical performance, with clinician validation.

  • Problem

    Whether clinical concept centers exist in these models and how to locate them remains the central question; rigorous clinical audits need access to internal model information.

    Source: §1 · p. 2

  • Method

    The authors use expert-assigned ICD codes to create a labeled EHR corpus for dictionary learning, then study concept centers across eleven open-weight Qwen and Gemma models.

    Source: §1 · pp. 2–3

  • Results

    Centers were recovered in all 11 models; steering yielded +1.1 to +2.3 additional correct diagnoses per hundred cases on the Qwen models, and activation gaps showed 76% agreement with clinician preferences.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Clinical-model evaluation can include mechanistic latents alongside generated text, and concept representations can serve as steerable vectors.

    Source: §1 · p. 2, §8 · p. 10

    The clinician-preference analysis hypothesizes that, within each pair, the response using the aligned concept center more is judged more meaningful.

    Source: §7 · p. 9

Abstract

from arXiv · show

Large language models are increasingly used in clinical settings. However, research into the reliability and performance of these models has focused almost entirely on the language substrate, scoring what the model says. Mechanistic interpretability has found that the latent space carries a higher fidelity of representation than the text: internal representations not only encode substantially more than the output verbalizes, but the stated reasoning also systematically omits features that causally drive the answer. An evaluation of model behavior in terms of mechanistic interpretability has not been explored in clinical decision support. In this work, we extend behavioral evaluation into the latent space and ask whether clinical concepts exist as locatable, causally used representations inside open-weight LLMs. We find dedicated clinical concept centers in the latent space of all eleven open models we test. These concept centers are interpretable, firing only on their aligned clinical narratives, and meaningfully and causally drive model behavior in both constrained and open-ended settings. They are not just analytical representations, but circuits that can be utilized in clinical practice, and we explore their use from the perspective of both evaluation and performance. From the evaluation standpoint, models stay internally coherent and keep using the relevant concept centers even under adversarial role-based priming, while aligned priming improves downstream clinical performance. From a performance perspective, we simulate realistic deployment settings and find that steering models along these centers leads to meaningful downstream improvements. Finally, we conduct a blinded clinician validation and find the activation and usage of these concept centers predicts clinicians preferences.

1 INTRODUCTION

Clinical LLM evaluation largely scores generated text, although internal representations can encode information and causal drivers that the text omits. This work uses expert-assigned ICD labels to locate clinical concept centers and tests their behavior across open-weight models, finding separable representations and downstream evaluation and performance uses.

1 INTRODUCTION

  • Clinical evaluations mainly score model outputs, while latent representations can encode correct answers and causal drivers omitted from generated text and stated reasoning.The authors argue output-level evaluations alone are insufficient to detect or prevent hidden latent-space shortcuts in clinical settings.Source: §1 · p. 2
  • The study uses clinical notes paired with expert-assigned ICD codes as a labeled corpus to locate clinical concepts in existing models’ latent spaces.This adapts established clinical labels for interpretability rather than training a model on the ontology.Source: §1 · p. 2
  • The authors establish concept-center existence, specificity, causal use, and persistence across eleven open-weight Qwen and Gemma models spanning 4B–30B parameters.Models include dense and mixture-of-experts, base, instruction-tuned, and medical variants.Source: §1 · p. 2
  • The ICD-label formulation provides expert-annotated supervision at scale without labeling effort and avoids automated-labeling error, with potential extension to domains whose records have established ontologies.The authors describe this as a clinically grounded formulation for latent-space interpretability.Source: §1 · p. 3
  • Held-out single-feature AUC reaches 0.80–0.92 in Qwen models, while the centers respond to clinical narratives and transfer across clinical settings.The reported transfer retains 45–86% of excess-over-chance separability.Source: §1 · p. 3
  • Aligned priming improves accuracy by +2.9 to +11.2 points, steering adds +1.1 to +2.3 correct diagnoses per hundred Qwen cases, and blinded clinicians’ preferences agree with the largest activation gaps 76% of the time.The reported preference agreement holds regardless of prompting.Source: §1 · p. 3

2 RELATED WORK

Prior clinical LLM evaluation has focused on language outputs, while mechanistic interpretability grounded in expert diagnostic labels and tested against clinical behavior was absent. The paper positions its work amid both latent-interpretability methods and skepticism about whether SAE features yield useful downstream behavior.

2 RELATED WORK

  • Clinical evaluation and interpretability have largely analyzed language outputs, and prior latent work had not grounded recovered features in expert diagnostic labels and tested their causal effect on clinical behavior.The cited clinical literature spans exam-style question answering, physician-rubric grading, and several early interpretability efforts.Source: §2 · p. 3
  • SAE usefulness is contested, so the authors retain probes and raw directions as baselines and report that the SAE-decoder direction is the stronger causal lever on their clinical task.They also use control-task logic and erasure methods to structure their evaluations.Source: §2 · p. 3

3 METHOD

The method derives clinical concept directions from discharge notes and expert ICD labels using sparse autoencoder features, then compares them with raw activation directions and probes. The evaluation corpus is split by patient to separate direction selection from reported held-out results.

3 METHOD

  • The study uses MIMIC-IV discharge summaries with expert-assigned ICD-9 codes, retaining primary diagnoses and grouping them into 17 disease chapters.After patient-level deduplication and capping at 1,000 notes per chapter, the corpus contains 8,617 notes; models of 27B parameters and above use a balanced 4,334-note subset.Source: §3 · p. 4
  • For each chapter, the authors difference mean SAE codes inside versus outside the chapter and decode that direction into the model’s residual-stream coordinates.They mean-pool note activations, encode them into sparse features, and use a patient-level selection/evaluation split.Source: §3 · p. 4
  • They compare the SAE direction with a raw difference-of-means direction and a multinomial logistic probe, using latent-direction projection as the basic measurement.The probe calibrates linear separability, while the SAE direction provides interpretable features and is reported as the stronger causal axis in later experiments.Source: §3 · p. 4

4 THE FOUNDATIONAL SETTING: CLINICAL CONCEPT CENTERS

The authors establish clinical concept centers as separable, clinically specific representations and test whether interventions affect model decisions and persist during generation. Their results show cross-model causal effects, a separation between where concepts are readable and where steering is effective, and sustained activation through long-form output.

4 THE FOUNDATIONAL SETTING: CLINICAL CONCEPT CENTERS

  • The foundational analysis treats decodability and causal impact as separate requirements for a reliable mechanistic account of clinical model behavior.The section investigates whether concept centers are distinctly separable and whether they shape outputs and behaviors.Source: §4 · pp. 4–5

4.1 DO CONCEPT CENTERS EXIST, AND ACROSS WHICH ARCHITECTURES AND SCALES?

  • One SAE feature reaches held-out AUC 0.80–0.92 in Qwen and 0.70–0.90 in Gemma, approaching the 0.90–0.93 macro-AUC logistic-probe ceiling on ten of eleven models.Separability is nearly uniform across the depth covered by the dictionaries; a stricter patient-level split leaves held-out AUC above 0.877.Source: §4.1 · p. 5, Figure 1 · p. 5
  • Concept directions generalize beyond discovery data, retaining 45–86% of excess-over-chance separability on patient-question text across Qwen models without refitting.A fresh SAE trained on II-Medical-8B recovers directions with average residual-space cosine +0.835 to the parent dictionary’s directions.Source: §4.1 · p. 5

4.2 ARE THE RECOVERED FEATURES SPECIFICALLY CLINICAL?

  • Features respond to contextual clinical narratives rather than merely diagnosis labels, discriminate adjacent chapters, and rarely activate on non-clinical text.For Qwen, activation falls from ten of ten features on full notes to 2.0–2.7 on bare diagnoses, 0.3–0.6 on adjacent-chapter notes, and at most 0.016 on non-clinical text.Source: §4.2 · p. 5

4.3 DOES THE MODEL CAUSALLY USE THE CONCEPT CENTERS FOR ITS DECISIONS?

  • The best intervention layer does not match the most separable layer in any of five Qwen models; effective steering instead falls around 60–85% of depth.On Qwen3-8B-Base, the behavioral swing in the effective band is 0.283 versus 0.022 at other depths with similar readout separability.Source: §4.3 · pp. 6–7
  • Held-out steering effects are positive across all 11 models, with no Qwen confidence interval touching zero and every Gemma model moving in the predicted direction.Matched random directions achieve far less; effects are bidirectional, largely orthogonal to answer unembedding rows, and show an inverted-U dose response.Source: §4.3 · p. 6, Table 1 · p. 7

4.4 DO CONCEPT CENTERS PERSIST THROUGH AUTOREGRESSIVE GENERATION?

  • Target-chapter dominance remains positive at every generated position on every model through the generation window, supporting per-token monitoring of persistent concept activity.Qwen3.5-27B sustains an effect size near 0.5 from token 100 to token 2,000; Gemma shows the same shape at smaller magnitudes.Source: §4.4 · p. 7

5 THE EVALUATION PERSPECTIVE: COUNTERFACTUAL ROLE PRIMING

The section tests whether specialist role prompts shift internal clinical concepts and diagnostic accuracy. Matched roles improve accuracy, while models remain internally coherent under counterfactual roles.

5 THE EVALUATION PERSPECTIVE: COUNTERFACTUAL ROLE PRIMING

  • Across 170 passes per model, Figure 2 tracks target-concept dominance per generated token, log-binned through 2,000 tokens, to assess persistence.The caption describes the measurement as an effect size.Source: Figure 2 · p. 8
  • The counterfactual design compares matched, counterfactual, random clinical, generalist, non-clinical, and no-role prompts using chapter projection and downstream accuracy.Projections are measured at each model’s best causal layer.Source: §5 · p. 8

5.1 DO INTERNAL CLINICAL CONCEPTS MATTER MORE THAN ROLE BASED PRIMING?

  • Matched-specialty priming improves accuracy by +2.9 to +11.2 points over the no-role baseline, while counterfactual priming performs almost similarly to baseline.The matched specialty raises projection onto the matching chapter direction; randomly drawn clinical specialties do not produce the same effect.Source: §5.1 · p. 8
  • The six-condition comparison attributes the shift to specialty identity: random clinical roles yield −0.08 projection shift (p = 0.83), versus +1.30 matched-minus-random (p = 2.3 × 10−19).The matched-minus-random result covers 14 of 17 chapters.Source: §5.1 · p. 8

6 THE PERFORMANCE PERSPECTIVE: MODEL STEERING

The section tests steering a model toward a clinically relevant concept center using deployment metadata. Steering along the true chapter direction improves open-ended diagnosis generation, while random or wrong directions do not help.

6 THE PERFORMANCE PERSPECTIVE: MODEL STEERING

  • The intervention uses deployment metadata, such as department or clinician specialty, to select a matching concept center before any text arrives.The section investigates whether upweighting that center improves the model’s answer.Source: §6 · p. 8
  • For open-ended generation, the intervention stays active at every generated token while each model writes a diagnosis for 1,624 held-out notes.The generated text is scored on whether it predicts the correct chapter.Source: §6 · pp. 8–9
  • True-chapter steering yields +1.1 to +2.3 additional correct diagnoses per hundred cases across three models; random-direction steering does not help, and wrong-chapter steering hurts.The result spans both model families and combines to p = 0.00046 using Fisher’s method.Source: §6 · p. 9

7 CAN LATENT CONCEPT CENTERS PREDICT CLINICIAN PREFERENCE?

A blinded clinician study compares responses generated under different role prompts and tests whether concept-center activation predicts preference. Greater separation in activation between responses corresponds to stronger agreement with clinician choices.

7 CAN LATENT CONCEPT CENTERS PREDICT CLINICIAN PREFERENCE?

  • Two clinical raters made 228 blinded, position-randomized head-to-head judgments, comparing responses to matched-specialty, adversarial other-specialty, and no-role prompts.The study used HealthBench cases matched to the Musculoskeletal chapter and recorded per-token concept traces and chapter projections.Source: §7 · p. 9
  • Target-chapter dominance is the average excess activation of the cohort’s chapter over all chapters while a response is being written.The per-token trace is reduced at the representation-peak layer to mean target-chapter dominance.Source: §7 · p. 9
  • Clinician agreement with activation-strength predictions rises from 56% across all pairs to 66% for the half with greater activation separation and 76% for the quarter with the greatest separation.Agreement remains at chance for aligned-versus-no-role comparisons, which have the smallest activation gaps.Source: §7 · p. 9
  • Role priming does not guarantee better open-ended output quality; similar concept-center use between responses offers no basis for choosing between them.The authors suggest using the mechanistic probe to guide abstention and selection for pass-at-k deployment.Source: §7 · p. 9

8 CONCLUSION

The conclusion presents clinical concept centers as locatable, specific, causally used, and persistent representations that support evaluation and performance studies. The appendices describe related work, tools, task setup, and model and dictionary specifications.

8 CONCLUSION

  • The study finds clinical concept centers are locatable, specific, causally used, persistent across generation, and more effective than the linguistic substrate.The authors argue that evaluation and use of medical models should include the latents in which they compute, alongside generated text.Source: §8 · p. 10
  • The authors state that no output is intended for clinical use without regulatory review.The work used credentialed-access MIMIC-IV data, public HealthBench conversations, and blinded judgments from two consenting clinical raters.Source: ETHICS STATEMENT · p. 10
  • The supplementary materials include experiment code and protocols, while MIMIC-IV requires credentialed access and is not redistributed.The paper specifies patient-level splits, model and dictionary identifiers, held-out selection, seeds, and statistical conventions for rebuilding or recomputing results.Source: REPRODUCIBILITY STATEMENT · p. 10, ETHICS STATEMENT · p. 10
  • An LLM-based coding assistant helped implement and execute experiments, compute reported statistics, and reformat notes; authors reviewed the work and take responsibility for it.The authors say they did not use generative AI to create synthetic datasets, formulate or prove mathematical claims, or propose hypotheses.Source: AI USE STATEMENT · p. 10

A EXTENDED RELATED WORK

  • The related-work discussion distinguishes this study’s expert-maintained clinical ontology and causal testing from prior SAE work using generated labels or correlational analysis.It also describes extending interpretability beyond specialized architectures and sub-domains to general-purpose Qwen and Gemma models, with clinical-task and clinician-preference applications.Source: A EXTENDED RELATED WORK · p. 15
  • Sparse-dictionary caveats motivate chapter-level concepts, dense raw-residual projections, and a dictionary-independence check; steering-vector reliability concerns motivate relative dosing and random controls.The authors identify SAE features as a possible finer-grained instrument for moving from chapters to smaller clinical concepts.Source: A EXTENDED RELATED WORK · p. 15

B COMPUTE AND TOOLKIT USED

  • Experiments used PyTorch and HuggingFace Transformers, with Gemma dictionaries loaded through SAELens and activation access using TransformerLens-style hooks.Runs used Nvidia H100 and A100 GPUs.Source: B COMPUTE AND TOOLKIT USED · p. 15

C PROMPT TEMPLATE AND CHAPTER MAPPING

  • The task prompt uses full, untruncated MIMIC-IV discharge summaries for ICD-9 chapter prediction; decoding uses temperature 0.0 and top-p 1.0.The chapter manifest contains 8,617 notes after per-chapter capping at 1,000; notes beyond context limits are truncated from the end.Source: C PROMPT TEMPLATE AND CHAPTER MAPPING · p. 16

E EXISTENCE: HELD-OUT SEPARABILITY AND SPLIT CONTROLS

Held-out and split-control results support cross-patient separability, while an independently trained dictionary recovers similar chapter directions. The supplied tables also identify the manifest and model-dictionary configurations used in the analysis.

E EXISTENCE: HELD-OUT SEPARABILITY AND SPLIT CONTROLS

  • The full manifest covers seventeen ICD-9-CM disease chapters, with Circulatory and Respiratory each capped at 1,000 notes and a majority-class baseline of 11.6%.The table lists the chapters, code ranges, and per-chapter sample sizes.Source: Table 3 · p. 17
  • Table 4 lists model and dictionary identifiers and states that MedGemma dictionaries are borrowed from the corresponding Gemma-3 parents.Source: Table 4 · p. 17
  • Held-out AUC exceeds 0.877 under patient-level splits, and in-sample versus held-out optimism stays below 0.009 in every case.Optimism is negative on four of five Qwen models, so held-out AUC often exceeds the in-sample value.Source: E EXISTENCE: HELD-OUT SEPARABILITY AND… · pp. 16–17, Table 5 · p. 17
  • An independently trained II-Medical-8B dictionary yields chapter directions with mean residual-space cosine +0.835 versus the parent dictionary, positive across all seventeen chapters.The cosine range is +0.784 to +0.865.Source: E EXISTENCE: HELD-OUT SEPARABILITY AND… · p. 17

F CLINICAL SPECIFICITY: FULL FIRING-RATE RESULTS

The firing-rate results compare activation of selected chapter features across conditions, with full discharge notes activating all ten selected features by construction. The same qualitative pattern is reported for the Gemma family.

F CLINICAL SPECIFICITY: FULL FIRING-RATE RESULTS

  • Full discharge notes activate all ten selected chapter features by construction because selection chose the ten highest-AUC features on discharge notes.Table 6 reports the Qwen family’s mean number of top-10 features activating per condition, averaged over seventeen chapters.Source: F CLINICAL SPECIFICITY: FULL… · p. 18, Table 6 · p. 18
  • A feature counts as active when it is among a token’s top-K features for TopK dictionaries or exceeds its learned pre-activation threshold for JumpReLU dictionaries.The table’s non-clinical-text condition uses fifteen short sentences about geography, sports, and food.Source: F CLINICAL SPECIFICITY: FULL… · p. 18, Table 6 · p. 18
  • The Gemma family follows the same qualitative condition-wise firing pattern as the Qwen family.Source: F CLINICAL SPECIFICITY: FULL… · p. 18

G CAUSAL INTERVENTION: SWEEP DETAIL AND CONTROLS

The intervention results report positive held-out effects and a depth-dependent response, alongside out-of-register transfer and sustained generation-time activation. The appendices also describe a borrowed-dictionary limitation for MedGemma and the measurement conditions for temporal traces.

G CAUSAL INTERVENTION: SWEEP DETAIL AND CONTROLS

  • Held-out ∆P is +0.041 to +0.069 for four Qwen models with sound baselines, while Gemma effects reach +0.0193.Matched-magnitude random perturbations are destructive on pretrained Gemma models, averaging −0.035 on 4B and −0.032 on 27B, while the chapter direction helps.Source: G CAUSAL INTERVENTION: SWEEP DETAIL… · p. 18
  • On Qwen3-8B-Base, earlier layers with similar separability swing 0.283 across injection strengths, versus 0.022 around the marginal AUC-maximizing layers.The sweep visual shows that positive response concentrates in a layer band despite nearly uniform separability across depth, and grows with strength up to the swept maximum.Source: G CAUSAL INTERVENTION: SWEEP DETAIL… · p. 18, Figure 3 · p. 19
  • The unembedding control fails to reproduce the steering effect, consistent with the near-zero cosine (+0.003) between its answer-letter vector and the chapter direction.Benjamini-Hochberg correction is applied across roughly 1,500 layer-strength combinations per model.Source: G CAUSAL INTERVENTION: SWEEP DETAIL… · p. 18
  • For MedGemma, dictionaries trained on Gemma-3 parents mean the specificity value reflects both the model’s medical training and how the borrowed dictionary interacts with its activations.Training dictionaries natively on these variants is left for future work.Source: G CAUSAL INTERVENTION: SWEEP DETAIL… · pp. 18–19

H REGISTER TRANSFER

  • Qwen chapter directions transfer without refitting to patient-question text, with every entry reaching permutation p < 1/2,000 under a 2,000-shuffle null.Retention is normalized by the in-corpus ceiling because AUC has a 0.5 floor.Source: H REGISTER TRANSFER · p. 19, Table 7 · p. 20

I TEMPORAL PERSISTENCE: MEASUREMENT DETAIL

  • Temporal traces use ten held-out notes per chapter, and base and chat models show the same sustained dominance shape during generation.Generations end naturally or at 2,000 tokens; most models exceed 1,200 tokens at the median, while II-Medical-8B has a 359-token median. Models are read at a responsive layer with at least 95% of the best layer’s finite-token coverage; II-Medical-8B’s projection rises later in the window.Source: I TEMPORAL PERSISTENCE: MEASUREMENT… · pp. 19–20

J ROLE PRIMING: SIX-CONDITION DESIGN AND THE DEGRADED-MODEL

Role-priming contrasts test whether specialty identity moves clinical concept projections, while steering experiments examine model rescue, steering-axis choice, and dose calibration. The results show distinct depth and template effects, with intervention outcomes depending on the direction and model-relative dose.

J ROLE PRIMING: SIX-CONDITION DESIGN AND THE DEGRADED-MODEL

  • Matched versus random-clinical priming raises true-chapter projections by +1.3010 on Qwen3-8B-Base, while counterfactual priming stays near the no-role baseline.The matched-over-random contrast exceeds t = 7.6 on four of five models; accuracy gains hold on 15–16 of 17 chapters per Qwen model.Source: J ROLE PRIMING: SIX-CONDITION DESIGN… · p. 20, Table 8 · p. 20
  • Persona-effect and chapter-encoding peaks move in opposite directions with model scale, so a fixed-depth readout misses one signal on most models.On Qwen3.5-27B, the persona-specific mean shift is 3.5× larger without the chat template, which dilutes the signal at the default final-position read.Source: J ROLE PRIMING: SIX-CONDITION DESIGN… · p. 21
  • II-Medical-8B’s correct-answer probability rises from 0.062 to 0.135 at the optimum when its chapter direction is injected at the causal-depth layers.The constrained-letter task scores 6.2%, despite strong internal separation: mean projection +74 for the true chapter versus −4.9 for wrong chapters.Source: J ROLE PRIMING: SIX-CONDITION DESIGN… · p. 21

K SAE-DECODER AGAINST RAW DIFFERENCE-OF-MEANS AS THE STEERING

  • On Qwen3-8B-Base, the SAE-decoder direction peaks at +0.038 while the raw direction produces no positive dose; on Qwen3-30B-A3B, the raw direction is destructive at every tested positive dose.On II-Medical-8B, both directions rescue output, but the SAE-decoder direction is about 70% more effective.Source: K SAE-DECODER AGAINST RAW… · p. 21, Table 9 · p. 22

L DOSE CALIBRATION FOR GENERATION STEERING

  • On Gemma-3-4B-it, relative dose ρ = 0.55 yields +0.0123 and 118 discordant answers; an absolute Qwen-sized dose changes 13 answers versus 12 for random control.The absolute dose is only 0.02× the natural between-chapter contrast on Gemma-3-4B-it.Source: L DOSE CALIBRATION FOR GENERATION… · p. 21, Table 10 · p. 22
  • Steering doses are scaled to each model’s median chapter-contrast norm, and discordant-answer counts distinguish an intervention that never fired from a true null.The SAE-decoder axis is preferred because it concentrates the concept on feature axes downstream computation reads, unlike correlational components in the raw mean direction.Source: L DOSE CALIBRATION FOR GENERATION… · p. 22

M CLINICIAN STUDY PROTOCOL

The blinded clinician study tests whether concept-center activation tracks response preferences across role-priming conditions. The discussion describes clinical concepts as depth-spanning circuits whose causal influence varies by layer.

M CLINICIAN STUDY PROTOCOL

  • The study compared Qwen3.5-27B responses under no-role, matched-specialty, and adversarial roles, using randomized blinded judgments and frozen activation-based pair predictions.The orthopedic cohort was chapter-faithful; raters judged 41 cases, and response identities were verified against stored generations.Source: M CLINICIAN STUDY PROTOCOL · pp. 22–23
  • Three literature-dependent cases were excluded because absent web access made reference verification impossible, so their ratings measured verification limits rather than response quality.The authors released the exclusion list, curated workbooks, and analysis code.Source: M CLINICIAN STUDY PROTOCOL · p. 23
  • Clinicians showed moderate agreement (raw 0.58; Cohen’s kappa 0.43), while activation-prediction agreement was 57.9% for the junior and 53.5% for the senior.The raters disagreed on 48 of 114 paired comparisons, often weighting completeness differently from technical correctness and context-seeking.Source: M CLINICIAN STUDY PROTOCOL · pp. 23–24
  • On non-tie judgments, plain responses beat adversarial priming 50–20 and matched-specialty priming 41–23, so role prompts did not reliably predict open-ended quality.Concept-center activations tracked quality where they separated responses.Source: M CLINICIAN STUDY PROTOCOL · pp. 23–24
  • 76.0% (38 of 50, p = 3.1 × 10−4) pooled agreement with clinician preferences occurred on pairs with the largest activation gaps.Agreement was 64.3% overall and 77.8% among larger-gap plain-versus-adversarial pairs, but at chance for aligned-versus-plain pairs.Source: M CLINICIAN STUDY PROTOCOL · p. 23, Table 11 · p. 24
  • Role prompts sometimes introduced unsupported specifics, specialty-framing intrusion, or overconfident calibration in open-ended responses.The qualitative examples contrast these effects with closed-task gains, where priming moves a classification answer rather than the framing and details clinicians assess.Source: M CLINICIAN STUDY PROTOCOL · pp. 24–25

N EXTENDED DISCUSSION

  • Clinical concept representations span model depth, but only some layers are causally potent; the persona effect occurs later and shifts later with scale.The discussion reports the persona effect at 83% of depth in the 8B model and 97% in the 27B, and distinguishes targeted impulses for constrained answers from sustained steering for open-ended generation.Source: N EXTENDED DISCUSSION · p. 25

Related papers

Loading 2610.02829v2…