PaperlayerView exact revision on arXiv

Source-linked AI summary

J-Miner: Recovering Executable Decision Knowledge from Language-Model Classifiers

Yunfan Gao, Xinyi Huang, Tao Sheng, Haorui Song, Yun Xiong, Haofen Wang

arXiv:2608.17063v1cs.LGcs.CL

TL;DR

Fine-tuned language-model classifiers encode task-relevant decision knowledge internally while exposing only final labels, limiting its inspectability and reuse. J-Miner recovers named concepts and executable rules, achieving 80.4% mean teacher fidelity with one concept and outperforming capacity-matched surface rules.

  • Problem

    Fine-tuned language-model classifiers embed task-relevant semantic patterns and decision regularities internally rather than exposing them through their output labels.

  • Method

    J-Miner maps internal classifier states into vocabulary space, extracts named concepts, and induces compact executable Boolean and weighted rules over them.

  • Results

    80.4% mean teacher fidelity is achieved across five tasks with one selected concept and an intercept, while compact concept sets consistently outperform matched surface features.

  • Takeaways & Limitations

    J-Miner makes classifier decision knowledge executable, inspectable, and transferable to compact standalone students that reconstruct concepts from raw text.

  • Takeaways & Limitations

    Recovered fidelity depends on how hidden states are measured and represented, including the chosen internal basis and calibration distribution.

Abstract

from arXiv · show

Large language models can be fine-tuned into specialized classifiers that perform well across diverse text tasks and make complex judgments, but they typically expose only final labels, leaving the decision knowledge acquired through fine-tuning implicit within the model. We study how to mine this internal decision knowledge from a fine-tuned classifier and encode it in an executable representation that can be inspected, validated, and reused beyond the source classifier. We introduce J-Miner, which mines text-level named concepts by aggregating vocabulary-aligned internal signals across layers and token positions, and uses the classifier's own predictions to learn executable decision rules over them. This process distills local internal readouts into an explicit classifier-level knowledge representation. Across multiple classification tasks, J-Miner rules reproduce up to 98.3\% of source-classifier decisions and achieve 6.0--29.5 percentage points higher behavioral fidelity than equally compact rules learned from input words. Further analysis shows that the named concepts reflect internal semantic evidence associated with task decisions, while the learned rules consolidate these distributed signals into inspectable decision structures. The resulting decision knowledge also transfers to lightweight standalone students: using about 1/24 as many parameters as the source classifiers, they reconstruct and execute the representation from raw text while retaining 99.8\% of the source classifiers' mean task accuracy. These findings show that task-specific decision knowledge can be faithfully represented in an explicit, executable form and reused beyond the classifier in which it was learned.

1 Introduction

J-Miner addresses the problem that fine-tuned LLM classifiers embed task-relevant decision knowledge without exposing it for inspection or reuse. It recovers vocabulary-named internal concepts and organizes them into compact executable rules that can be validated and transferred to lightweight students.

  • Fine-tuning embeds task-relevant semantic patterns and decision regularities inside LLM classifiers, making them difficult to observe, explain, or reuse from output labels alone.
  • J-Miner maps intermediate classifier states into vocabulary space, recovers vocabulary-named internal concepts, and mines compact Boolean rules and weighted decision logic.The recovered representation connects interpretation of classifier knowledge with downstream reuse.
  • Across diverse binary and multiclass tasks, J-Miner identifies small sets of internal concepts whose names, weights, and per-example contributions expose decision evidence.Concepts frequently activate without literal occurrence of their names.
  • 33.2M E5-small students reproduce 88.7–98.0% of source-classifier decisions directly from raw text while executing fixed decision logic without the source classifier at inference.
  • J-Miner produces decision knowledge that can be examined, behaviorally validated, and reused beyond the source classifier.

2 Related Work

Prior work has adapted pretrained language models for varied text-classification tasks, but their deployed interfaces typically expose only predicted labels. Knowledge extraction and rule-extraction methods make model behavior more explicit, increasingly grounding decision logic in neural representations and causally tested internal units.

  • Language-model classifiers: Pretrained language models support text classification through classification heads or verbalized labels, including toxicity, abusive-language, and hate-speech detection.Their deployed interface usually exposes only the predicted label.
  • Knowledge and rule extraction: Knowledge extraction seeks an understandable model-level account of how isolated evidence or features combine into predictions.Rule extraction is one approach to making this combination explicit.
  • Knowledge and rule extraction: Classical rule-extraction methods approximate black-box predictors with trees or symbolic expressions, emphasizing global rules and behavioral fidelity.Their predicates are usually defined over inputs or introduced outside the model.
  • Neural representation grounding: Recent methods move rule extraction closer to neural representations by grounding predicates in causally tested neurons or extracting logic over thresholded internal units.MechaRule uses causally tested neurons, while NEUROLOGIC extracts logic over thresholded internal units.

3 Problem Definition

J-Miner treats a frozen language-model classifier as a source of internal decision knowledge, using read-only internal states and verdict logits to recover named variables and an executable prediction rule. The goal is a compact vocabulary of concepts, their per-input activations, and a rule that combines them into the classifier’s verdicts.

  • Source classifier: J-Miner assumes a frozen classifier and read-only access to its internal states and verdict logits.The classifier predicts classes represented by verdict tokens, with z_θ(y|x) denoting the logit of verdict token y.
  • Recovered representation: For concept budget K, J-Miner names K internal coordinates with tokens from the classifier’s vocabulary.These coordinates form C_K = {v_1, ..., v_K}.
  • Recovered representation: J-Miner represents each input with a K-dimensional binary activation state and learns a rule g: {0, 1}^K → Y over these concepts.The activation state is denoted c_θ(x), while g maps concept activations to class verdicts.
  • Executable decision knowledge: The resulting representation consists of named concepts, per-input activations, and an executable rule specifying how concepts combine into a verdict.Together, these components define the compact decision representation sought by J-Miner.

4 Method

J-Miner recovers vocabulary-aligned internal readouts as recurring message-level concepts, fits executable decision rules over them, reveals their contributions, and transfers the fixed representation to a compact student model. Its rules use either Boolean structures or signed scorecards, while student inference reconstructs concepts from text without the source classifier or J-Lens.

  • Pipeline: J-Miner’s Recover, Reveal, and Transfer stages respectively mine concepts and rules, expose decision contributions, and pair the fixed rule with student-predicted concepts.Recover uses J-Lens readouts; Reveal analyzes coefficients and example-level contributions; Transfer trains a compact text model to reconstruct concepts.
  • Recover: J-Lens maps intermediate residual states into rankings over the classifier’s vocabulary, whose recurring coordinates become message-level concepts after aggregation across layers and positions.The aggregation distinguishes concepts from single-token explanations by recording whether a vocabulary coordinate appears anywhere in the designated message region.
  • Recover: Classifier predictions select concepts by one-vs-rest trigger-rate gaps after insufficiently supported coordinates are removed, while dataset labels do not guide ranking or rule fitting.The selected concept vector has capacity K, and prediction-derived verdicts provide the construction target.
  • Rule fitting: The executable representation combines size-bounded Boolean ASTs with a signed linear scorecard whose coefficients, intercept, threshold, and concept vocabulary specify how evidence accumulates.Boolean rules express conjunction, disjunction, and absence; the scorecard preserves a direct mapping from each weighted term to a named concept.
  • Transfer: Transfer trains a compact student reader on J-Lens-derived supervision to predict binary concept inputs, then applies the unchanged executable rule after probability thresholding.The source classifier and J-Lens are unnecessary at inference, and the rule receives no gradient or refitting during student training.

5 Experimental Setup

J-Miner is evaluated across nine English classification tasks using independently fine-tuned, frozen Qwen3.5-0.8B classifiers. Experiments compare matched compact-rule and interpretability baselines, control Transfer conditions, and measure task performance, behavioral agreement, and internal evidence properties.

  • Task and classifier setup: J-Miner is evaluated on nine English text-classification tasks—six binary and three multiclass—with independently fine-tuned Qwen3.5-0.8B classifiers frozen before application.Recover and Reveal cover all nine tasks; Transfer covers the six binary tasks and extends to SNIPS-3 and SNIPS-7.
  • Compact-rule comparisons: The matched Surface baseline ranks literal input words against teacher predictions, selects the same feature budget K, and fits the same class-balanced LR-Kheads.Both routes use L2 regularization with C=1.0, lbfgs, at most 2,000 iterations, and identical frozen test examples.
  • Interpretability baselines: Interpretability comparisons include vanilla readout, SAE, attention rollout, Integrated Gradients, and a 1,024-dimensional dense probe using shared construction and test conditions.The methods differ in whether they extract vocabulary coordinates, sparse latent features, token importance, or classifier-verdict predictions.
  • Transfer setup: Transfer isolates standalone-prediction sources by fixing the 33.2M E5-small-v2 encoder, training data, budget, and test examples across Surface, J-only, and J+Aux.Surface predicts literal indicators, J-only predicts J-Miner concepts, and J+Aux adds a verdict loss to concept reconstruction.
  • Evaluation metrics: Evaluation reports accuracy, teacher fidelity, and multiclass macro-F1, while Reveal additionally measures non-literal activation, scorecard concentration, and layer-wise label-state stabilization.Accuracy measures dataset-label prediction quality, whereas teacher fidelity measures agreement with the corresponding frozen classifier.

6 Results

J-Miner recovers compact executable rules that closely reproduce classifier decisions across binary and multiclass tasks, model scales, and classifier families. The named concept–rule interface also supports interpretable decision analysis and transfers effectively to lightweight standalone students.

  • Executable recovery: LR-16 reaches .750–.983 teacher fidelity across six binary tasks, improving over AST-5 by 0.33–6.17 percentage points on every task.AST-5 uses five signed concept literals, whereas LR-16 weights all 16 selected concepts.
  • Concept basis: 6.00–29.50 percentage points are the gains in teacher fidelity of LR-16 over matched Surface scorecards, with a 17.08-point unweighted mean gain.Every task-level paired interval lies above zero.
  • Multiclass recovery: On HateXplain, SNIPS-3, and SNIPS-7, AST-5 reaches .673, .987, and .865 teacher fidelity, respectively, in one-versus-rest multiclass ensembles.LR-16 learns one thresholded score for each class, while AST-5 learns one five-literal rule per class.
  • Scale and model families: .773–.953 teacher fidelity is achieved across 24 matched task–classifier cells spanning three Qwen sizes and Gemma, Llama, and Phi classifiers.The recovered rules exceed matched Surface rules by 11.5–35.5 percentage points.
  • Interpretability: Decision concentration requires a task-level median of 3 to 11 active concepts to cover 90% of absolute feature contribution, with Sarcasm concentrating in three concepts.Sentiment, Formality, IMDB, and Toxicity require four to six concepts.
  • Standalone transfer: 91.8% mean gold accuracy and 92.4% mean fidelity are achieved by the 33.2M J+Aux student, compared with 92.0% source accuracy, while reproducing 88.7%–98.0% of source decisions.No task differs from the source classifiers by more than 1.67 percentage points.

7 Discussion

J-Miner shows that task-specific decision behavior in evaluated LLM classifiers is highly compressible, with compact concept rules recovering substantial teacher behavior. Its explicit concept states and fixed rules separate decision knowledge from source-model parameters, while fidelity depends on hidden-state measurement and representation.

  • Compression: 80.4% mean teacher fidelity is reached across five tasks with one selected concept and an intercept, while compact concept sets recover more decisions.The matched advantage over surface features recurs across tested model sizes and families.
  • Measurement and representation: Fidelity depends on hidden-state measurement and representation: J-Lens uses task-agnostic external text to fit a common vocabulary-space readout.Average Jacobian transport and vocabulary projection create an information bottleneck, and the alternative-basis comparison exposes this design choice.
  • Portability: J-Miner makes decision knowledge portable by having a compact student reconstruct the recovered concept state and execute the fixed rule outside the source classifier.The named decision variables and their composition remain available and comparable as the implementing model changes.

8 Conclusion · A Frozen concept–rule construction

J-Miner recovers executable decision knowledge by mining named concepts from decision-relevant internal readouts and composing them into rules. Its frozen concept–rule object uses message-level concepts and a fitted scorecard, while a capacity-matched literal-word route provides the comparison.

  • 8 Conclusion: J-Miner mines named concepts from decision-relevant internal readouts and composes them into executable rules.The method targets executable decision knowledge from language-model classifiers.
  • 8 Conclusion: The resulting representation reproduces source-classifier decisions more faithfully than capacity-matched literal-word rules.The comparison uses rules based on concepts versus literal input words.
  • 8 Conclusion: Non-literal concept activations help make the recovered decision structure directly inspectable.The passage identifies non-literal concept activations as part of the inspection mechanism.
  • 8 Conclusion: Per-example concept contributions further expose the recovered decision structure for direct inspection.These contributions complement the named concepts and executable rules.
  • A Frozen concept–rule construction: For each task, the frozen concept–rule object is constructed from recurring J-Lens readouts.Its components are selected per task rather than described as a single universal object.
  • A Frozen concept–rule construction: The object contains a selected message-level vocabulary, an indicator map, and a fitted scorecard.These are the three explicitly listed components of the concept–rule object.
  • A Frozen concept–rule construction: The capacity-matched Surface object instead derives its indicators from literal words in the input.Both objects are paired for comparison, with indicators differing in source.
  • A Frozen concept–rule construction: The concept and Surface routes share teacher targets, feature budgets, scorecard families, and held-out examples.They differ only in the source of their indicators.

A.1 Data separation and object freezing … B Extended validation of executable recovery

J-Miner recovers executable teacher decision knowledge by freezing data-derived objects, aggregating internal readouts into message-level concepts, and comparing them with capacity-matched surface features. Extended validation examines which interface factors preserve teacher behavior across the recovery pipeline.

  • A.1 Data separation and object freezing: J-Miner separates lens fitting, rule construction, and held-out evaluation across distinct data roles, reserving gold labels for accuracy readback.External calibration text fits the averaged Jacobian lens, while frozen teacher predictions guide construction.
  • A.1 Data separation and object freezing: After construction, the selected vocabulary, indicator map, and fitted scorecard remain fixed, reducing comparison to their evidence-source transformations.The matched Surface route produces an analogous object within the same data boundaries.
  • A.2 Message-level indicator construction: J-Lens readouts are pooled across content positions and source layers into message-level indicators, excluding the instruction template and terminal Decision: position.Primary Qwen classifiers retain the three highest-ranked vocabulary coordinates at each content position across layers 14–22 and the final-layer identity readout at layer 23.
  • A.2 Message-level indicator construction: Each retained vocabulary coordinate becomes an executable whole-message variable when it appears anywhere in the retained readouts.This aggregation converts position- and depth-dependent coordinates into message-level indicators.
  • A.2 Message-level indicator construction: 16 coordinates form the frozen J-Miner vocabulary after construction-only support filtering and ranking by teacher-prediction trigger-rate differences.Coordinates occurring in fewer than .03 of construction examples are discarded before ranking.
  • A.3 Capacity-matched Surface construction: The capacity-matched Surface route builds a 16-word vocabulary from lowercased raw text, using distinct words of at least two characters that appear in at least .03 of construction examples.Candidates are split on non-alphanumeric characters other than apostrophes.
  • A.3 Capacity-matched Surface construction: Both indicator sets fit the same class-balanced logistic scorecard under fixed construction, teacher-target, feature-budget, weighting, model-family, and test-example conditions.Binary tasks use a zero threshold, while multiclass tasks select the highest-scoring class.
  • B Extended validation of executable recovery: Extended validation separates late-layer label-state access from concept compression and tests whether the recovery direction recurs across model scales and additional conditions.The primary recovery result spans six binary tasks and favors message-level concepts from J-Lens readouts over capacity-matched Surface features.

B.1 Binary classification tasks statistics

Across six binary classification tasks, J-Miner makes late-layer teacher decisions nearly readable and compresses them into sixteen executable concepts while preserving more teacher fidelity than input-word rules. It also captures teacher-specific behavior, including many decisions that disagree with gold labels.

  • Binary classification tasks statistics: Label-state fidelity ranges from .991 to .998 across six tasks, showing that selected late layers nearly expose the teacher verdict.The label state uses the decision-position J-Lens signals from late layers.
  • Binary classification tasks statistics: .750–.983 teacher fidelity is retained when late-layer readouts are compressed into sixteen recurring message-level concepts.The compression target is the teacher’s decision rather than the dataset label.
  • Binary classification tasks statistics: 6.00 to 29.50 percentage points is J-Miner’s fidelity advantage over the capacity-matched Surface scorecard, with every task-level paired interval above zero.The comparison is between executable rules over internal concepts and rules over literal input words.
  • Binary classification tasks statistics: Five of six tasks show J-Miner reproducing a larger share of teacher errors than Surface, with SMS the only reversal among six examples.Teacher-error fidelity is evaluated only on held-out examples where the frozen teacher disagrees with the dataset label.

B.2 Results across model scales and families

Across model scales and families, J-Miner consistently recovers teacher decisions more faithfully than capacity-matched Surface scorecards. This internal-over-Surface advantage holds across every evaluated task–classifier setting, though gap magnitude varies by checkpoint and task.

  • Qwen family: J-Miner outperforms Surface in all twelve Qwen task–classifier cells from 0.8B to 4B.Recovered fidelity and gap magnitude vary across checkpoints, with task-dependent scale trajectories.
  • Qwen family: 20.92 percentage points is the task-equal mean gap for Qwen 2B, versus 19.42 points for Qwen 4B.Formality has the smallest gap at each Qwen scale, while Toxicity has the largest.
  • Other model families: All twelve Gemma, Llama, and Phi cells satisfy F_J > F_S, with mean gaps of 20.67, 18.54, and 19.08 percentage points, respectively.The corresponding 95% intervals are [18.63, 22.71], [16.42, 20.58], and [17.00, 21.13], all above zero.
  • Overall comparison: All 24 matched cells across the full panel favor J-Miner over Surface.The experiments span Qwen, Gemma, Llama, and Phi classifiers on Sentiment, Formality, IMDB, and Toxicity, using K=16 matched scorecards.

B.3 Feature budgets and candidate ranking … C.2 Composition and label-state stabilization

Across budgets, tasks, and representations, J-Miner recovers teacher decisions through compact internal-concept scorecards whose performance depends on candidate ranking, feature capacity, and class coverage. The recovered interface exposes non-literal concept activations, compositional signed contributions, and layer-wise stabilization of label states.

  • B.3 Feature budgets and candidate ranking: B.3 Feature budgets and candidate ranking: .804 mean fidelity with one J-Miner concept versus .614 for one Surface word, while all 30 budget gaps remain positive from K=1 to K=32.Across five tasks, the gaps range from +4.67 to +33.33 percentage points.
  • B.3 Feature budgets and candidate ranking: B.3 Feature budgets and candidate ranking: Teacher-ranked orderings exceed random-order medians in every displayed task-by-budget cell, with SMS reaching .980 fidelity using four concepts versus a .668 random median.The comparison uses 100 random nested orderings from the same internal candidate pool.
  • B.4 Multiclass recovery and validation: B.4 Multiclass recovery and validation: SNIPS-3 represents all three classes with sixteen concepts and retains .993/.993 J-Miner teacher fidelity, whereas SNIPS-7 represents five of seven classes.SNIPS-7 has .998/.998 label-state fidelity, but 126 examples fall back to the intercept-defined prediction because classes 5 and 6 lack selected coordinates.
  • C. Inspecting the recovered interface: C. Inspecting the recovered interface: The executable scorecard maps named internal coordinates to signed weighted contributions plus an intercept, exposing how task-specific evidence composes each verdict.The persisted LR-16 objects retain all sixteen coordinates, coefficients, and intercepts, while displayed terms show the strongest signed directions.
  • C.1 Non-literal activation of named coordinates: C.1 Non-literal activation of named coordinates: All-coordinate mean non-literal activation ranges from .8045 to .9861, remaining between .7766 and .9798 when restricted to names observed in held-out text.Thus, most activations in the observed-name subset occur without the corresponding printed word form.
  • C.2 Composition and label-state stabilization: C.2 Composition and label-state stabilization: The minimum support for 90% of absolute contribution varies by task, with Sarcasm concentrated in three coordinates, four to six needed for several tasks, and SMS requiring eleven.The scorecard’s example-level contributions are paired with layer traces whose L95 occurs between L10 and L15 across all six binary tasks.
  • C.2 Composition and label-state stabilization: C.2 Composition and label-state stabilization: Label states stabilize at L14 for Formality and Sarcasm and L15 for Sentiment, IMDB, and Toxicity, while class-specific median settling layers differ by up to twelve layers.In the SMS trace, the label state settles on spam from L14 as position-level readouts develop from price and discount into coupon and promo.

C.3 Alternative feature representations · C.4 Multiclass coverage under a fixed budget

J-Miner matches alternative representations in compact predictive fidelity while exposing directly named, executable vocabulary predicates. Under a fixed multiclass budget, the scorecard’s retained vocabulary determines which class directions remain explicitly representable, despite stable classifier label-state readouts.

  • C.3 Alternative feature representations: Across four tasks, SAE has the highest fidelity on Toxicity, Formality, and IMDB, while J-Miner reaches .8200 on Sentiment versus SAE at .8083.The comparison places J-Miner and SAE in a similar compact-fidelity range while preserving vocabulary coordinates in an explicit scorecard.
  • C.3 Alternative feature representations: J-Miner’s matched Toxicity rules reach truth accuracies of .832 and .845 while directly composing absolutely, angry, hate, and viol.The named predicates retain their vocabulary identity inside the executable rule.
  • C.3 Alternative feature representations: J-Miner preserves predicate identity and execution: selected-name CWS ranges from .7500 to 1.0000, versus .0000–.2269 for SAE, without external lookup.SAE rules require four activation-token lookup rows, whereas J-Miner composes named vocabulary predicates directly.
  • C.3 Alternative feature representations: Predictive naming remains near chance: .5066–.5668 for J-Miner, .5003–.5589 for vanilla readout, and .5251–.5640 for SAE.Dictionary-coordinate count and literal-name predictive naming measure different properties; name occurrence does not separate the internal representations.
  • C.4 Multiclass coverage under a fixed budget: Label-state fidelities reach .937/.921 for HateXplain, 1.000/1.000 for SNIPS-3, and .998/.998 for SNIPS-7, reported as accuracy/macro-F1.The multiclass label state is the argmax over class label-token readouts; L95 occurs at L21, L15, and L15, respectively.
  • C.4 Multiclass coverage under a fixed budget: Among active-concept examples, HateXplain, SNIPS-3, and SNIPS-7 require medians of three, four, and three coordinates to cover 90% of absolute winner–runner-up contribution.Zero-feature rows number 134/600, 2/297, and 126/665, where intercepts determine predictions.
  • C.4 Multiclass coverage under a fixed budget: SNIPS-3 retains all three class directions, reaches .993/.993 teacher fidelity, and has minimum class recall .980; SNIPS-7 covers only five of seven classes.SNIPS-3 has two zero-feature test examples, while SNIPS-7 lacks selected coordinates for classes 5 and 6.
  • C.4 Multiclass coverage under a fixed budget: The fixed budget controls both rule size and explicit class-space coverage: stable label-token verdicts coexist with selective retention of class directions in the executable scorecard.SNIPS-3 demonstrates complete three-class coverage, whereas SNIPS-7 demonstrates incomplete seven-class coverage despite high label-state fidelity.
Loading 2608.17063v1…