PaperlayerView exact revision on arXiv

Source-linked AI summary

Concentrated Causal Influence, Distributed Diagnostic Information: Cognitive Impairment Decisions in Large Audio-Language Models

Xiutian Zhao, Luqi Sun, Björn Schuller, Berrak Sisman

arXiv:2610.32061v1eess.AScs.CL

TL;DR

The paper asks which internal components control LALMs’ CI decisions from speech and how diagnostic information is distributed in their representations. It identifies and intervenes on CI-contrast neurons across inputs and corpora, finding concentrated causal control alongside broadly available diagnostic information.

  • Problem

    LALM answers do not reveal which internal computations transform speech and transcript evidence into CI decisions or how diagnostic information is represented.

    Source: §1 · p. 1

  • Method

    The study selects class-contrastive decoder MLP neurons and tests their causal influence against matched controls across input conditions and corpora.

    Source: §1 · p. 2

  • Results

    In three of five eligible models, CICN ablation produces class-directed decision shifts beyond matched controls, while diagnostic information is also available in larger neuron sets.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    The findings distinguish concentrated causal control over output decisions from broadly distributed diagnostic information in LALMs.

    Source: §6 · p. 8

    These experiments do not validate LALMs as diagnostic tools, and picture-description performance does not establish clinical validity or reliability across populations.

    Source: Ethical Considerations · p. 9

Abstract

from arXiv · show

Large audio-language models (LALMs) can assess cognitive impairment (CI) from speech, yet their answers reveal neither which internal components control the decision nor how much CI information is accessible in their representations. We investigate both questions through CI-contrast neurons (CICNs), selected by class-contrastive activation frequency in decoder MLPs. Across eight open LALMs and two corpora (Pitt and PROCESS-2), we causally test CICNs against size- and layer-matched random masks and activity-matched controls. In three of the five models that meet our eligibility criteria, ablation of small CICN sets causes large, class-directed prediction shifts; speech-identified masks also control transcript-only decisions, and three of four primary Pitt masks retain specificity on PROCESS-2. Position-restricted interventions locate the influence before the final answer position, most strongly after the speech or transcript evidence under letter-labeled options, largely as a decision-direction shift. Beyond causal localization, supervised linear readouts of frozen CICN activations achieve 0.92--0.97 AUC within Pitt, versus 0.67--0.84 for the models' direct answers. Matched random and all-neuron readouts reach comparable performance; top-ranked CICNs show only an exploratory advantage at very small unit budgets. These results distinguish concentrated causal control over output decisions from broadly distributed diagnostic information for CI detection in LALMs.

1 Introduction

The paper asks how LALMs transform speech and transcript evidence into cognitive-impairment decisions, distinguishing internal causal control from information available in representations. It identifies CI-contrast neurons and tests their effects across inputs and corpora.

Motivation

  • Prior output-accuracy measures cannot reveal which evidence sources models use or which internal computations control their CI decisions.Speech can provide timing, expressiveness, and disfluency information beyond transcripts, while benchmark corpora may also contain shortcuts.Source: §1 · p. 1

Approach

  • CICNs are decoder MLP neurons selected by class-contrastive activation-frequency differences between correctly and confidently classified healthy and impaired recordings.The study evaluates eight open LALMs on Pitt and PROCESS-2 across speech-only, transcript-only, and joint inputs.Source: §1 · p. 2

Main findings

  • In three of five eligible models, ablating small CI-contrast neuron sets causes class-directed shifts beyond matched controls, with effects transferring across input modalities and, for several masks, corpora.The strongest localized effects occur before the final answer position, especially after content under letter-labeled options; readout evidence is more distributed.Source: §1 · p. 2

Contributions

  • The study introduces a causal neuron-level analysis of CI decisions in LALMs, testing specificity with matched random and activity-matched controls and examining transfer, position, and information distribution.These analyses distinguish concentrated causal influence from diagnostic information accessible across a broader set of neurons.Source: §1 · p. 2

2 Related Work

Prior work studies neuron selectivity and input-level explanations, but causal localization of CI decisions in models that process speech directly remains unexplored in the cited literature.

Neuron-Level Interpretability

  • Earlier neuron-level multimodal work mainly examined modality- or domain-specific units, and the cited closest CI-related neuron study used a text-only LLM.The paper positions causal localization in speech-processing models and shared control across speech and transcript inputs as open questions.Source: §2 · pp. 2–3
  • Multimodal neuron studies have emphasized modality- or domain-specific units, while whether selected units are privileged predictive features remains unclear.Sparse-autoencoder features and selected-neuron fine-tuning have produced results ranging from parity with comparison features to gains over random-neuron baselines.Source: §2 · p. 2

Interpreting CI Detection Models

  • CI-detector interpretability has mainly attributed predictions to input features or transcript tokens, rather than locating the internal computations producing them.The cited methods include feature importance, saliency, attention, and SHAP.Source: §2 · p. 3

3 Methods

The methods define a binary CI probe, identify class-contrastive decoder MLP neurons from confident correct decisions, and test their effects against matched controls. Additional analyses examine input conditions, intervention positions, linear readouts, and statistical criteria.

3.1 Probe Task

  • The probe asks whether a speaker is cognitively healthy or impaired, records decoder MLP activations on correct predictions, and tests selected masks against size- and layer-matched random controls.Both option orders are used to counterbalance letter and position preferences.Source: §3 · p. 3, §3.1 · p. 3

3.2 Input Conditions

  • Four conditions vary the available evidence: transcript-only, speech-only, joint speech with clean transcript, and joint speech with annotated transcript.The annotated transcript adds fillers, fragments, pauses, retracings, and unintelligible speech in readable form; investigator prompts are excluded to avoid label leakage.Source: §3.2 · p. 3

3.3 Activation Statistics and CICN Identification

  • CICNs are the globally top-ranked 0.5% of decoder MLP neurons by class-contrastive activation-frequency differences, profiled using balanced pools of confident correct recordings.Activation frequency measures the fraction of positions with positive gate activation; the statistic selects neurons, while intervention tests functional importance.Source: §3.3 · pp. 3–4

3.4 Interventions

  • Ablation zeros the masked MLP coordinates, predicting movement away from the mask’s class: removing NC neurons should raise P(CI), while removing CI neurons should lower it.Interventions ordinarily apply at every prefill position.Source: §3.4 · p. 4

3.5 Specificity Baselines

  • Specificity is assessed with per-layer-count-matched random masks and, for selected Pitt masks, activity-matched sets drawn without class labels.The activity-matched sets also match activation probability and mean gate activation per layer.Source: §3.5 · p. 4

3.6 Where CICNs Act

  • Position-restricted ablations test audio and transcript spans, the final prompt position, non-final positions, and template positions before or after content.Each variant is compared with random masks restricted in the same way.Source: §3.6 · p. 4

3.7 Exploitation Tests

  • Linear-readout exploitation tests compare mean-activation ridge readouts from CICNs with model answers, matched random neurons, and an all-neuron reference.Fixed feature budgets from 8 to 2,048 units compare top-ranked CICNs with layer-matched random neurons; the unit-budget analysis excludes recordings used for mask selection.Source: §3.7 · p. 4

3.8 Metrics and Statistical Testing

  • Intervention inference compares each mask with matched random masks on both ΔP(CI) and mean signed per-item log-odds change, requiring both metrics to pass corrected significance testing.The intersection–union test uses the larger of the two p-values, with Benjamini–Hochberg correction; flipown is descriptive and not part of significance testing.Source: §3.8 · pp. 4–5

4 Experiment Setup

The experiments cover eight open LALMs and two Cookie Theft corpora, with identification restricted to model-condition cells meeting predefined baseline and pool-size criteria.

Models

  • Eight open LALMs receive the same 16 kHz mono waveform, and their audio-length limits do not bind in these experiments.The models span different zero-shot operating regimes.Source: §4 · p. 5

Datasets

  • Pitt is the primary identification corpus, while PROCESS-2 provides an independent cross-corpus transfer setting with different collection and transcription regimes.MCI and other diagnoses are held out for transfer analyses, and PROCESS-2 transcript style is normalized and stratified because it is associated with diagnosis.Source: §4 · p. 5

Baseline performance and analysis eligibility

  • Eligibility criteria yield twelve Pitt model-condition cells from five models and two PROCESS-2 transcript cells for identification.Criteria require transcript UAR of at least 0.65, audio-condition bootstrap intervals excluding 0.5, and at least 15 qualifying identification recordings per class.Source: §4 · p. 5

5 Results

Small CICN ablations shift prompted CI decisions beyond matched controls in eligible models, with control transferring across inputs and corpora and localizing mainly before the final answer. Diagnostic information is more broadly distributed across neurons.

5.1 Concentrated Causal Control

  • Ablating 0.5% of decoder MLP neurons produced class-directed shifts beyond matched random masks in three models; seven of 24 masks passed both co-primary metrics, and four had |∆ℓ| = 1.6–2.6 per item versus |∆ℓ| ≤0.17 for matched random masks.All seven passing masks shifted in the predicted direction; the four largest-effect masks became the primary masks for follow-up analyses.Source: §5.1 · p. 5, Figure 1 · p. 6

5.2 Decision Bias versus Diagnostic Discrimination

  • For eight masks, 79–100% of recordings moved in the predicted direction versus 37–60% for random masks; a single offset explained 61–91% of primary-mask log-odds change while item ordering largely persisted.Each mask reordered items more than its random controls, but only two primary masks reduced AUC more than matched masks, and those changes were not significant against zero after correction.Source: §5.2 · p. 6, Table 1 · p. 6

5.3 Control Transfers across Modalities and Corpora

  • All nine transcript-only transfers passed against matched random masks; three of four primary masks also passed on PROCESS-2, and held-out diagnoses generally shifted in the same direction.Speech-identified masks affected transcript-only decisions, while Pitt-identified masks transferred to PROCESS-2, including recordings with held-out diagnoses such as MCI.Source: §5.3 · p. 7, Table 2 · p. 7

5.4 Localizing the Intervention Effects

  • Ablating only the final prompt position had little effect, while earlier positions approximately reproduced the full effect; the strongest influence followed the content and weakened under numeric option labels.The strongest localized effects arose at the question and options after the content, with smaller effects from audio and transcript spans; influence remained before the final answer position.Source: §5.4 · pp. 7–8, Figure 2 · p. 7

5.5 Distributed Diagnostic Information

  • Within Pitt, whole-mask CICN readouts reached 0.92–0.97 AUC versus 0.67–0.84 for model answers, but matched-random and all-neuron readouts approached the same performance.At eight units, top-ranked CICNs exceeded random draws by 0.118 AUC on average; the small-budget advantage was exploratory, and random features largely caught up by 128–512 units.Source: §5.5 · p. 8, Figure 3 · p. 8

6 Conclusion

The study distinguishes concentrated causal control over CI decisions from broadly distributed diagnostic information. Across eligible models, CICN ablations shift decisions across modalities and corpora, while diagnostic ranking is largely preserved; this distinction provides a basis for auditing LALM decisions.

6 Conclusion

  • In three of five eligible models, CICN ablation produced class-directed shifts beyond matched controls, and the same neuron sets influenced decisions across input modalities and corpora.The interventions primarily changed decision bias while largely preserving diagnostic ranking.Source: §6 · p. 8
  • The strongest localized effects occurred at post-content prompt positions before the final answer under letter-labeled options.Top-ranked CICNs showed an exploratory readout advantage at very small feature budgets, while larger random feature sets recovered comparable information.Source: §6 · p. 8
  • The distinction between concentrated causal control and distributed diagnostic information provides a basis for auditing how LALMs translate internal representations into diagnostic decisions.Source: §6 · p. 8

Limitations

The study’s inferences are bounded by its classifier, intervention targets, decision interface, and evaluated data.

Limitations

  • CICNs are defined from the model’s own confident decisions, so causal claims concern the prompted classifier; token-level activity links are correlations, not biomarker interpretations or causal evidence about those tokens.Duration, silence, and speech rate explain part of the selection contrast, while 58–91% of the class gap remains.Source: Limitations · p. 9
  • Interventions target decoder MLP neurons, leaving attention, the audio encoder, and multimodal adapters for future circuit-level tracing.The analysis uses twelve Pitt cells from five models, one identification prompt template, two answer options, first-step option logits, and linear readouts.Source: Limitations · p. 9
  • The study uses English Cookie Theft descriptions from two corpora with different recording and transcription regimes; PROCESS-2 evaluation includes 20 CI and 136 held-out MCI recordings.PROCESS-2 serves as an independent transfer environment.Source: Limitations · p. 9

Ethical Considerations

The authors treat cognitive-status recordings as sensitive research data and limit their use to aggregate analysis. They also state that the study does not establish clinical validity or justify deployment without further safeguards.

Ethical Considerations

  • The study uses access-controlled speech data only for research, reports aggregate results, and does not identify participants or infer beyond provided labels.Source: Ethical Considerations · p. 9
  • Picture-description performance does not establish clinical validity or reliability across populations, and cognitive impairment spans conditions from MCI to dementia.Source: Ethical Considerations · p. 9
  • Any deployment would require prospective clinical validation, demographic and recording-condition assessment, consent and privacy protections, human oversight, and regulatory compliance.Source: Ethical Considerations · p. 9

A.1 Models

  • The models receive common 16 kHz mono audio, with input handling adjusted for models whose limits would otherwise truncate recordings.Qwen3-Omni’s intended limit is restored to 300 seconds; Gemma audio is divided into clips of at most 30 seconds.Source: §A.1 · pp. 13–14

A.2 Implementation Details

The implementation combines participant-disjoint data preparation and controlled prompts with deterministic inference and mask-level statistical tests. Eligibility, transcript style, and recording conditions constrain how the resulting comparisons should be interpreted.

A.2 Implementation Details

  • Three of five eligible models produce class-directed CICN ablation shifts beyond matched controls, and their significance survives a broader eligibility correction.The four primary masks remain below q = 0.05 under the correction over all 32 Pitt model-condition cells.Source: §A.2 · pp. 14–15

A.3 Data Preparation

  • Pitt uses timed participant speech segments and excludes seven low-bandwidth 8 kHz recordings that could make audio bandwidth a label shortcut.PROCESS-2 is independently collected and lacks transcript timings, so its full recordings are used.Source: §A.3 · p. 15
  • PROCESS-2 transcript style is associated with diagnosis, and rendering and stratification reduce but cannot fully exclude this confound, especially for audio.PROCESS-2 recordings have no timings, so speech cannot be cut to participant utterances; automatic transcripts were not evaluated.Source: §A.3 · p. 16

A.4 Transcript Renderings

  • Pitt’s annotated transcript retains fillers and fragments and renders pauses and repair phenomena as text, while clean transcripts remove CHAT codes.PROCESS-2 clean transcripts remove parenthetical notes and fillers; its two transcript styles are also retained for analysis.Source: §A.4 · p. 16

A.5 Prompts

  • Each item is prompted with equally likely healthy and impaired options, and every item is run with both option orders to counterbalance letter and position preferences.Source: §A.5 · p. 17

A.6 Deterministic Inference

  • Predictions use first-step logits before decoding processors, and inference is greedy and deterministic under fixed seeds and hardware.Independent runs produced identical logits, predictions, and generated text on the same GPU type.Source: §A.6 · p. 17

B Statistical Procedure

  • Specificity is tested across matched masks rather than recordings, because item-level paired tests mark 75–95% of random masks significant.The random-mask distribution is tested with a prediction-interval t-test; the authors also report descriptive rank and bootstrap checks.Source: B Statistical Procedure · pp. 17–18

C.1 Baseline Performance

  • Adding audio significantly lowers PROCESS-2 UAR for Gemma-4-12B and Voxtral-Small-24B, while shifting their class priors in opposite directions.On Pitt, Qwen2.5-Omni-7B also changes substantially as transcript and rendered disfluency information are added; non-participant speech and measured recording-quality variables do not account for PROCESS-2 shifts.Source: §C.1 · p. 18

C.2 Recording Properties and the Effect of Audio

Recording properties are associated with some audio-induced decision shifts, but duration alone does not explain the PROCESS-2 modality effect. Robustness analyses find that CICN effects often survive stronger controls and prompt changes, while their magnitude depends on mask and identification-pool choices.

C.2 Recording Properties and the Effect of Audio

  • PROCESS-2 recordings are generally clean and show no bandwidth shortcut like the excluded Pitt recordings, though transcription styles differ slightly in cutoff frequency.Duration remains the most consistent recording-property correlate of the audio-induced shift on PROCESS-2.Source: §C.2 · pp. 18–19
  • Duration is the most consistently associated recording property with audio-induced shifts on PROCESS-2, but the modality effect persists in both duration halves.Associations are weaker on Pitt; neither non-participant speech nor class differences in recording length account for the shifts.Source: §C.2 · p. 19

C.3 Identification Pools

  • Table 13 reports eligible recordings per class and pooled tokens for cells reaching identification, but the supplied passage gives no cell values.The identification-pool criterion is documented separately from the reported pool counts.Source: §C.3 · p. 19

D.2 Activity-Matched and Alternative Selection Controls

  • Six of eight strongest Pitt masks, including all four primary masks, pass activity-matched controls in the predicted direction (q = 0.040).Kimi-Audio A NC does not pass; for Qwen3-Omni-30B AT CI, most activity-matched sets produce larger shifts in the opposite direction.Source: §D.2 · pp. 19, 21
  • The recording-property comparison is descriptive because random masks are matched to CICN masks, not separately to each property mask.Diagnosis contrast explains most selected-neuron identity, but recording-property masks can produce stronger effects in some models and the CICN shift direction is not unique to the CI contrast in every model.Source: §D.2 · p. 20

D.3 Sensitivity to Mask Size and Pool Size

  • Across 32 sensitivity cells, 29 preserve the main-effect direction; smaller masks and identification pools generally weaken effects, while larger masks often exceed the 0.5% setting.For Kimi-Audio AT-chat NC, the effect falls from +0.267 to +0.064 when the nested mask grows from 0.5% to 1%, showing that nested-mask effects are not additive.Source: §D.3 · p. 21, Table 15 · p. 21

D.4 Class Contrasts Adjusted for Recording Properties

  • After adjustment for duration, silence, and speech rate, CICN masks retain 58–91% of their class gap, with 97–100% of selected neurons keeping its direction in the pool.Across all identification recordings, 93–100% of selected neurons retain the direction of their class gap.Source: §D.4 · p. 22, Table 16 · p. 22

D.5 Prompt Variants

  • All four primary masks retain predicted-direction effects under both prompt variants, with magnitudes comparable to the original prompt.The variants reword the task or replace letter labels with numeric labels; Kimi-Audio AT-chat is unreliable under the paraphrase, so this demonstrates decision control rather than improvement of a reliable classifier.Source: §D.5 · p. 22

D.6 Speaker-Attribute Control: Full Results

Cross-task controls show that CICN effects are mostly selective to cognitive-impairment decisions, though some cross-effects occur. Steering the selected neurons can shift decisions, while ablation effects are dominated by an offset that leaves most thresholded decisions unchanged after correction.

D.6 Speaker-Attribute Control: Full Results

  • Six of seven CICN masks leave gender decisions unchanged, and five also keep gender log-odds within the matched-random range.Qwen3-Omni-30B AT CI is the main substantive exception; specificity is assessed only against gender, and gender decisions are near-certain compared with CI decisions.Source: §D.6 · p. 23, Table 17 · p. 23

D.7 Steering: Full Results

  • Shifting Kimi-Audio and MiniCPM-o-4.5 neurons along their class-mean differences steers decisions monotonically and reversibly with α; Qwen2.5-Omni-7B responds little.No mask–α cell survives family correction against norm-matched random pushes, despite descriptive directional control in several Kimi-Audio cells.Source: §D.7 · p. 23, Table 18 · p. 23

E Decision Bias versus Diagnostic Discrimination

  • After subtracting the mean intervention shift, 90–97% of primary-mask thresholded decisions match the unmodified model, while AUC is unchanged by that correction.The two AUC decreases exceed matched-random decreases, but paired bootstrap intervals for ∆AUC against zero span zero across the eight masks.Source: E Decision Bias versus Diagnostic… · p. 24

F.1 Cross-Modal Transfer, Mask Overlap, and Span Controls

CICN interventions transfer across input conditions and, for several masks, across corpora, but their effects depend on the mask and evaluation setting. Span and held-out tests qualify the scope of that transfer.

F.1 Cross-Modal Transfer, Mask Overlap, and Span Controls

  • Mask depth profiles are broadly consistent across conditions within models, but differ across models and include a corpus-dependent shift for Gemma-4-12B.Kimi-Audio, MiniCPM-o-4.5, and Qwen3-Omni-30B are comparatively front-loaded; Qwen2.5-Omni-7B and Gemma-4-12B place little mask mass in their earliest layers.Source: Figure 4 · p. 24, §F.1 · p. 25
  • Audio-span NC masks from Kimi-Audio and Qwen3-Omni-30B shift transcript-only decisions about as much as or more than speech-only decisions, while text-span effects are weaker.Token-matched audio masks largely preserve the causal behavior of full audio-span masks; span selection does not always guarantee the predicted class direction.Source: §F.1 · p. 25
  • All three primary joint-condition masks retain effects comparable to matched-transcript settings when transcripts are replaced with ones from the opposite class, without identifying how evidence is combined.The result supports robustness to conflicting inputs, not a claim about how the model combines them.Source: §F.1 · p. 25

F.2 Steering Transcript-Only Decisions

  • Kimi-Audio’s audio-discovered NC masks steer transcript-only decisions in eight of nine tested mask–α cells, with one A-mask exception.Specificity is clearer under transcript-only input because matched-random steering effects vary less.Source: §F.2 · p. 25

F.3 Cross-Corpus Transfer to PROCESS-2

  • On PROCESS-2, four Pitt-passing masks also pass, and two additional Pitt-nonsignificant masks pass; transferred effects indicate shared decision shifts, not shared evidence use.No effect is confined to one transcription style; PROCESS-2’s NC skew leaves too few correctly classified CI items for stable CI flip rates, so inference uses ΔP(CI) and Δℓ across recordings.Source: §F.3 · pp. 25–26

F.4 Held-Out Diagnoses

  • Held-out tests pass for seven of eight Pitt masks, five of six PROCESS-2 masks, and both transferred AT-chat masks; Qwen3-Omni-30B’s AT CI mask is the exception.Pitt effects remain similar when restricted to held-out MCI cases, so they are not driven by the small number of non-MCI diagnoses.Source: §F.4 · p. 26, Table 23 · p. 26

G.1 Position-Restricted Ablation

Position-restricted interventions place CICN influence before the final answer position, with localization depending on prompt format. Token-level analyses associate mask activity with picture content and disfluency cues, while a depth readout places most CICNs before answer-level diagnosticity emerges.

G.1 Position-Restricted Ablation

  • Transcript-position ablations of four masks produce 11–38% of the corresponding all-position Δℓ; both speech-only masks exceed matched random controls on both co-primary metrics.These are separate interventions, not additive mediation shares; position-restricted localization covers Kimi-Audio and MiniCPM-o-4.5 only because the implementation does not support the relevant sparse mixture-of-experts intervention.Source: §G.1 · p. 26, Table 24 · p. 27

G.2 Positional Robustness across Prompt Variants

  • After-content positions carry 53–79% of all-but-last Δℓ under paraphrased prompts, but only 24–56% under numeric labels, where no mask passes.Influence before the final position holds across prompts; concentration at the question and options is supported for letter-labeled options and weakens with numeric labels.Source: §G.2 · p. 27

G.3 Answer-Aligned Readout across Depth

  • 77–92% of CICNs lie before the depth where the residual-stream option readout reaches answer-level AUC, consistent with an upstream role for these neurons.The readout remains near or below chance through much of the network; it does not establish when diagnostic information first becomes internally available.Source: §G.3 · pp. 27–28, Figure 5 · p. 28

G.4 Token-Level Correlates of CICN Activity

  • CICN masks are more NC-like on picture-content words and more CI-like on fillers, fragments, pauses, retracings, and silence beyond random masks, with one exception.Removing marker, silence, and information-unit tokens leaves 95–101% of the class gap, indicating a recording-level state carried by ordinary tokens; these associations do not establish cue mediation of the dominant decision shift.Source: §G.4 · p. 28, Table 25 · p. 28

H Distributed Diagnostic Information

Frozen-activation readouts show strong within-Pitt CI decodability, but comparable performance from matched-random and all-neuron features indicates that diagnostic information is not concentrated in CICNs. Cross-corpus results and small-budget comparisons qualify that conclusion.

H Distributed Diagnostic Information

  • On Pitt, supervised CICN readouts achieved 0.92–0.97 AUC versus 0.67–0.84 for the model’s own answers, outperforming them in all twelve within-corpus cells.Their AUC was in the range of the clean-transcript TF-IDF baseline (0.91); differences were not tested.Source: Figure 7 · p. 29, H Distributed Diagnostic Information · pp. 29–30
  • On PROCESS-2, Pitt-trained CICN readouts reach 0.76–0.82 AUC and exceed model answers in eight of twelve cells, without a consistent advantage over matched-random or all-neuron readouts.Source: Figure 7 · p. 29, H Distributed Diagnostic Information · p. 30
  • Projecting out linearly decodable recording-quality and transcription-style confounds changes CI-readout AUC by −0.033 to +0.014 across the reported settings.The confound directions are estimated after class means are removed; the authors conclude the CI readouts depend little on the linearly decodable part of either confound.Source: H Distributed Diagnostic Information · p. 30
  • At 8 and 32 units, top-ranked CICNs outperform mean random draws in all seven cells by 0.04–0.18 and 0.003–0.12 AUC, respectively, but the advantage is exploratory.No individual budget comparison survives correction in the reported training set; tiny random subsets are highly variable.Source: H Distributed Diagnostic Information · p. 30
  • All-neuron transcript-only readouts reach 0.93 AUC for Qwen2.5-Omni-7B and 0.95 for Qwen3-Omni-30B, versus model-answer AUCs of 0.75 and 0.78.These cells have no CICN masks, so the result does not depend on an audio shortcut.Source: H Distributed Diagnostic Information · p. 30

I License

  • The study uses Pitt and PROCESS-2 under their terms, redistributes neither audio nor transcripts, and reports only aggregate results.Source: I License · p. 30

Related papers

Loading 2610.32061v1…