J-Miner: Recovering the Decision Logic of Fine-Tuned LLM Classifiers as Compact Rules

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

cs.LGAug 17, 2026

J-Miner: Recovering the Decision Logic of Fine-Tuned LLM Classifiers as Compact Rules

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

Organizations: Shanghai Research Institute for Intelligent Autonomous Systems, Tongji University · Shanghai Key Laboratory of Data Science, College of Computer Science and Artificial Intelligence, Fudan University · Meituan · College of Design and Innovation, Tongji University

Abstract

Task-fine-tuned large language model (LLM) classifiers acquire task-specific decision knowledge, but this knowledge remains implicit in distributed internal computations, making their decision logic difficult to interpret. We introduce the Executable Decision Compression (EDC) framework and propose J-Miner, which mines vocabulary-named variables from internal readouts and learns rules shared across inputs to produce executable explanations. Analysis reveals that a small set of these variables captures much of the classifier's decision behavior, holding for both varying parameter scales within a family and distinct families. Across six binary tasks, a rule using just one variable reproduces 76.7% of source-classifier decisions on average, rising to 88.8% with 16 variables. Most of the decision information retained by these variables comes from internal activations beyond literal surface matching. A lightweight text reader predicts the variable states, allowing the same fixed rules to execute independently of the source classifier.

Figures & tables

Figure 1: A fine-tuned LLM classifier outputs only a label, leaving decision logic implicit. J-Miner fits a shared executable rule over vocabulary-named internal readouts to reproduce its predictions.

Figure 2: Overview of Executable Decision Compression (EDC) and J-Miner.

Figure 3: Compact rule recovery and layer-wise label readout. (a) Held-out fidelity to the source classifier, averaged over five tasks. (b) Held-out J-Lens verdict agreement with final predictions.

		Binary						Multiclass
Method	Metric (%)	SMS	Sentiment	Formality	IMDB	Toxicity	Sarcasm	HateXplain	SNIPS-3	SNIPS-7
Source	Acc.	98.0	90.5	94.8	94.5	85.5	88.8	62.2	100.0	98.2
Zero-shot	Acc.	51.3	66.8	54.3	76.2	60.7	50.0	35.2	69.4	74.6
Majority	Fid.	50.0	56.2	51.5	52.2	57.2	54.8	47.7	33.3	14.0
Surface	Fid.	92.3	62.3	78.7	69.8	61.0	66.3	65.3	98.0	81.7
AST	Acc.	96.7	80.7	91.2	82.5	84.3	68.8	51.2	98.7	85.1

Table 1: Recovering classifier decisions with compact rules.

Figure 4: Recovering and executing decision rules. (a) Source fidelity (pale: Surface). (b) Boolean and linear replay. (c) Reader replacement (diamonds: net change). Mean weights tasks equally.

	Source-classifier fidelity (%)				Concept F1		Cover (%)		Keep (%)
Task	$g(c_{\theta})$	J-only	J+Aux	Plain KD	J-only	J+Aux	J-only	J+Aux	J-only	J+Aux
SMS	98.3	97.3	98.7	98.3	.919	.914	98.0	98.3	98.3	98.3
Sentiment	84.3	90.2	92.0	91.3	.571	.564	79.8	81.0	82.8	83.3
Formality	95.3	92.3	95.5	95.5	.738	.696	89.3	93.5	92.7	93.8
IMDB	89.5	92.7	94.2	93.2	.738	.721	86.2	88.2	90.5	89.3
Toxicity	90.5	89.0	89.8	88.5	.682	.672	88.7	88.3	90.2	89.7

Table 2: Standalone rule execution on text-predicted concept states.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

		Exhaustive $K{=}1$		Exhaustive $K{=}2$		Prefix test
Task	Evidence	Constr.	Test	Constr.	Test	$K{=}1$	$K{=}2$	$K^{\star}_{0.10}$
SMS	J-Lens	96.7	95.0	97.7	96.7	95.0	96.7	1
	Literal	69.3	73.3	79.1	81.7	73.3	80.7	$\geq 3$
Sentiment	J-Lens	72.2	70.7	78.8	77.5	69.8	75.8	$\geq 3$
	Literal	55.5	55.3	57.8	58.3	50.0	52.2	$\geq 3$
Formality	J-Lens	85.3	82.8	89.8	87.3	83.3	85.3	3

Appendix

Table 3: Small-budget linear-rule fidelity (%) and certified minimum budgets (

K^{\star}_{0.10}

) across the six binary tasks.

Task	Train / Val. / Test	$\|D_{c}\|$	Epochs	Batch $\times$ Accum.	Max Tokens
SMS	896 / 298 / 300	1,194	4	$8\times 2$	192
Sentiment	1,800 / 600 / 600	2,400	4	$8\times 2$	256
Formality	1,800 / 600 / 600	2,400	4	$8\times 2$	256
IMDB	1,800 / 600 / 600	2,400	4	$4\times 4$	512
Toxicity	1,800 / 600 / 600	2,400	4	$8\times 2$	256
Sarcasm	1,800 / 600 / 600	2,400	4	$8\times 2$	256

Appendix

Table 4: Dataset partitions, construction sample sizes (

|D_{c}|

), and fine-tuning schedules for Qwen3.5-0.8B source classifiers.

Classifier	Tasks	Transported layers	Identity
Qwen3.5-0.8B	Binary	14–22	23
Qwen3.5-2B	Binary	14–22	23
Qwen3.5-4B	Binary	19–30	31
Gemma-3-1B	Binary	16–24	25
Llama-3.2-1B	Binary	10–14	15
Phi-4-mini	Binary	20–30	31

Appendix

Table 5: Fixed zero-based layer windows for message-level variable extraction.

Task	Task phrase	Labels (1 / 0)	Field	Additional excluded names
SMS	SMS spam	spam / ham	Message	spam, ham, sms, message, classification
Sentiment	sentiment	positive / negative	Text	sentiment, text
Formality	formality	formal / informal	Sentence	formal, informal, formality, sentence
IMDB	sentiment	positive / negative	Review	sentiment, review
Toxicity	toxicity	toxic / non-toxic	Comment	toxic, toxicity, non, comment, classification
Sarcasm	sarcasm	sarcastic / not sarcastic	Headline	sarcastic, sarcasm, not, headline

Appendix

Table 6: Prompt template fields and task-specific excluded vocabulary for the six binary classification tasks.

Task	$F_{J}$ (%)	$F_{S}$ (%)	$\Delta F$ (pp)	Paired 95% CI (pp)	McNemar $p$
SMS	98.3	92.3	+6.00	[3.00, 9.33]	$1.21\!\times\!10^{-4}$
Sentiment	84.3	62.3	+22.00	[17.50, 26.50]	$4.37\!\times\!10^{-20}$
Formality	95.3	78.7	+16.67	[13.33, 20.17]	$4.52\!\times\!10^{-21}$
IMDB	89.5	69.8	+19.67	[15.50, 24.00]	$2.24\!\times\!10^{-18}$
Toxicity	90.5	61.0	+29.50	[25.33, 33.67]	$3.14\!\times\!10^{-39}$
Sarcasm	75.0	66.3	+8.67	[3.50, 13.67]	$1.27\!\times\!10^{-3}$

Appendix

Table 7: Binary recovery fidelity (%) and paired statistical tests on the held-out test sets.

Task	$F^{J}_{\mathrm{gold}}$ (%)	Error count $n$	$F^{J}_{\mathrm{err}}$ (%)	$F^{J_{\mathrm{gold}}}_{\mathrm{err}}$ (%)	$F^{S}_{\mathrm{err}}$ (%)
SMS	98.0	6	50.0	50.0	66.7
Sentiment	84.7	57	68.4	63.2	49.1
Formality	95.5	31	77.4	77.4	58.1
IMDB	88.8	33	63.6	54.5	54.5
Toxicity	90.0	87	71.3	69.0	58.6
Sarcasm	75.2	67	55.2	56.7	35.8

Appendix

Table 8: Gold-fitted control and error-following fidelity (%) on the held-out binary test sets.

Task	Classifier	$F_{J}$ (%)	$F_{S}$ (%)	$A_{J}$ (%)	$A_{S}$ (%)	$\Delta F$ (pp)	95% CI (pp)
Sentiment	Qwen3.5-0.8B	84.3	62.3	80.8	62.5	+22.00	[17.50, 26.50]
	Qwen3.5-2B	82.2	63.3	79.2	64.0	+18.83	[14.50, 23.17]
	Qwen3.5-4B	78.0	64.2	78.8	63.7	+13.83	[9.17, 18.50]
	Gemma-3-1B	87.5	63.2	83.0	64.0	+24.33	[20.00, 28.83]
	Llama-3.2-1B	77.3	62.8	76.2	63.0	+14.50	[9.67, 19.33]
	Phi-4-mini	82.5	66.0	80.7	64.8	+16.50	[11.83, 21.17]

Appendix

Table 9: Recovery fidelity and gold accuracy (%) across model scales and families at

K=16

Figure 5: Source-classifier fidelity across matched feature budgets. Each panel compares J-Miner and Surface scorecards from

K=1

K=32

under the same source-classifier target and prediction head. The dotted line marks the primary

K=16

setting.

Task	$K=1$	$K=2$	$K=4$	$K=8$	$K=16$	$K=32$	$K=64$
SMS	95.0 (55.3)	96.7 (59.3)	98.0 (66.8)	98.3 (78.3)	98.3 (88.3)	98.7 (93.5)	98.7 (96.3)
Sentiment	69.8 (55.5)	75.8 (56.3)	79.5 (55.9)	83.7 (58.7)	84.3 (63.9)	84.7 (69.4)	87.3 (75.0)
Formality	83.3 (52.7)	85.3 (54.8)	92.3 (59.9)	94.0 (66.7)	95.3 (76.2)	96.5 (83.8)	96.5 (88.3)
IMDB	76.0 (52.2)	79.0 (52.7)	81.7 (53.7)	86.3 (55.4)	89.5 (59.0)	90.2 (63.3)	91.8 (69.4)
Toxicity	77.8 (55.0)	86.0 (54.3)	88.8 (55.3)	88.5 (56.4)	90.5 (61.1)	90.7 (68.8)	90.8 (78.0)
Sarcasm	58.3 (48.7)	63.7 (52.7)	69.5 (53.1)	71.8 (56.8)	75.0 (61.3)	78.7 (66.5)	82.8 (71.7)

Appendix

Table 10: Held-out fidelity (%) of

\Delta(v)

-ranked prefixes and random orderings across budgets

K

Figure 6: Held-out fidelity grids

\widehat{F}_{\mathcal{P}}(\ell,K)

across the six binary tasks. Upper heatmaps report held-out source-classifier fidelity over cumulative cutoff layers

\ell\in\{14,\dots,23\}

and variable budgets

K\in\{1,\dots,32\}

, with cells at or above

90\%

fidelity outlined; lower curves plot budget slices at cutoffs L14, L18, and L23.

Task	Test $n$	L14 (%)	L23 (%)	Range across cutoffs (%)
SMS	300	96.7	98.3	96.7–99.0
Sentiment	600	83.0	84.3	81.5–86.3
Formality	600	93.3	95.3	93.3–95.7
IMDB	600	87.7	89.5	86.5–89.5
Toxicity	600	89.0	90.5	89.0–90.8
Sarcasm	600	72.3	75.0	72.3–76.7

Appendix

Table 11: Held-out source-classifier fidelity (%) at

K=16

across observation depth cutoffs.

	Share (%)		Held-out fidelity (%)
Task	Non-literal gap	Literal activations	All	Literal	Non-literal	Surface
SMS	94.8	5.8	98.3	76.7	98.7	92.3
Sentiment	95.6	4.1	84.3	49.3	84.3	62.3
Formality	72.6	27.0	95.3	83.5	94.0	78.7
IMDB	91.3	6.6	89.5	65.7	88.5	69.8
Toxicity	99.1	1.5	90.5	45.8	90.7	61.0

Appendix

Table 12: Literal and non-literal activation shares (%) and refitted LR-16 held-out fidelity (%) across the six binary tasks.

Excluded positions	Sentiment (%)	Toxicity (%)
None, full content sequence	84.3	90.5
Sentence-ending punctuation only	82.5	85.3
Final content position	82.3	85.5
All tokens without letters or digits	80.5	84.7
Surface rule, full input	62.3	61.0

Appendix

Table 13: Held-out fidelity (%) of frozen LR-16 scorecards after excluding structural or boundary readout positions prior to pooling.

Figure 7: Recovered 16-variable linear scorecards for SMS, Sentiment, and Toxicity on Qwen3.5-0.8B, showing eight representative coefficients per task.

Figure 8: Direction-agreement gap between rule-guided and matched-random word deletions across the six binary classification tasks. Points report percentage-point gains in the rate at which edits shift the source classifier’s margin in the rule-predicted direction, with horizontal bars showing paired 95% confidence intervals and the vertical dashed line marking the six-task mean of

+12.45

pp. Absolute rates and eligible sample counts appear in Table 14 .

Task	Eligible $n/N$	Targeted	Random	Gap [95% CI]
		Agreement (%)
SMS	164/300	73.78	59.15	+14.63 [6.71, 22.56]
Sentiment	401/600	55.61	46.80	+8.81 [3.99, 13.63]
Formality	263/600	57.03	46.39	+10.65 [4.18, 17.11]
IMDB	482/600	51.45	39.00	+12.45 [8.44, 16.60]
Toxicity	219/600	59.36	37.90	+21.46 [12.79, 30.14]

Appendix

Table 14: Direction-agreement rates (%) and paired differences (pp) for rule-guided versus matched-random word deletions.

Task	J-Lens	Vanilla	SAE	Surface	J-Lens / Vanilla
	Observation protocol within EDC (%)			Input reference (%)	English words
Toxicity	90.50	91.83	93.33	61.00	16 / 3
Formality	95.33	96.33	97.83	78.67	15 / 10
Sentiment	84.33	86.17	81.67	62.33	16 / 6
IMDB	89.50	89.00	94.00	69.83	16 / 5
Mean / Total	89.92	90.83	91.71	67.96	63 / 24

Appendix

Table 15: Held-out source-classifier fidelity (%) and complete English word counts for 16-variable linear scorecards across internal observation protocols.

Task	Reader	$p_{\mathrm{corr}}$ (%)	$p_{\mathrm{harm}}$ (%)	$F_{\phi}-F_{\Pi}$ (pp)	Toward (%)
SMS	J-only	0.33	1.33	$-1.00$	91.3
	J+Aux	1.00	0.67	$+0.33$	94.9
Sentiment	J-only	11.50	5.67	$+5.83$	89.4
	J+Aux	12.17	4.50	$+7.67$	90.9
Formality	J-only	2.17	5.17	$-3.00$	88.1
	J+Aux	3.17	3.00	$+0.17$	94.9

Appendix

Table 16: Decision-change decomposition and directional alignment under standalone reader replacement.

Explore similar work

Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation
May 4, 2026Francesco Sovrano, Gabriele Dominici, Marc LangheinrichMechanistic Interpretability Ablation
Università della Svizzera italiana Lugano, Switzerland
Structured Prompt Optimization Meets Reinforcement Learning for Global and Local Interpretability over Complex Text
May 27, 2026Tianyang Zhou, Wenbo Chen, Pierre Jinghong Liang +1Large Language Model Classification Automatic Prompt Optimization
1Carnegie Mellon University · 2Amazon
LLM-as-Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them
Oct 1, 2026Yinheng Li, Justin WagleDecisions
Microsoft

May 4, 2026, cs.LGJ/K move · Enter open · S save

Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation

Francesco Sovrano, Gabriele Dominici, Marc Langheinrich

Università della Svizzera italiana Lugano, Switzerland

Mechanistic Interpretability Ablation Extraction

Open paper AskPDF arXiv

A central goal of explainable AI is to express large language model (LLM) decision logic symbolically and ground it in internal mechanisms. Existing rule-extraction methods usually learn ungrounded symbolic surrogates, while mechanistic interpretability links behavior to neurons but often requires hand-crafted hypotheses and costly interventions. We introduce MechaRule, a pipeline that grounds rule extraction in LLM circuits by localizing sparse agonist activations whose ablation disrupts rule-related behavior. MechaRule rests on two findings. First, in a fixed baseline/flip regime, sparse agonist effects can exhibit overtopping: a few high-effect activations remain detectable within larger groups, dominate weaker ones, and flip many of the same examples. In such regimes, adaptive group testing with confidence-guided conservative pruning requires O(k log(N/k) + k) interventions over N candidates when k << N are agonists. Second, agonists are localized more reliably on data splits aligned with close-to-faithful rule behavior; spectral splits provide a rule-free fallback, whereas unfaithful splits degrade localization. Empirically, on arithmetic and jailbreaking, MechaRule recalls 97.0% of highest-effect agonists in matched brute-force validations at only 2.14% of exhaustive-ablation cost on average. Ablating the localized agonists eliminates 97.6--100.0% of eligible correct arithmetic answers and jailbreaks, and can correct arithmetic errors or induce jailbreaks by up to 72.8% and 32.5%.

Abstract

Explore similar work

Neuron-Anchored Rule Extraction for Large Language Models via Contrastive Hierarchical Ablation

Structured Prompt Optimization Meets Reinforcement Learning for Global and Local Interpretability over Complex Text

LLM-as-Jev: LLMs Are Already Jev-Style Decision Models -- When and How to Fine-Tune Them