Agent-Guided Evidence Acquisition for Emotion-Cause Pair Extraction

Zhuangzhuang Pan; Yan Xia; Chee Seng Chan

cs.CLOct 10, 2026

Agent-Guided Evidence Acquisition for Emotion-Cause Pair Extraction

Authors: Zhuangzhuang Pan, Yan Xia, Chee Seng Chan

Organizations: Universiti Malaya, Malaysia · Suzhou University of Technology, China · VinUniversity, Vietnam

Abstract

Emotion-Cause Pair Extraction (ECPE) requires identifying the causes associated with emotions expressed in a conversation. Accurate extraction can depend on evidence that is not initially available, making it important to decide when further information is worth acquiring. We study agent-guided evidence acquisition for ECPE, where a student agent selectively accesses expert judgments for borderline relations on both sides of its validation-selected F1 decision boundary and revises only queried predictions. Through controlled comparisons under different information-access conditions, we examine how delegation granularity, query selection, and response integration affect prediction quality. At a nominal 10% query rate, selective fusion improves mean pair F1 by 2.64-3.88 points over the student in our main experiments, reaching 65.14 on ECF and 49.74 on MEC4 across model families. Our analysis suggests that greater access does not always translate into better cause identification. These findings motivate acquisition strategies that consider both the relevance of newly available evidence and its effect on prediction revision, providing an empirical basis for more effective and efficient evidence acquisition in ECPE.

Figures & tables

Figure 1. Query allocation with two event candidates. Each orange card uses one expert answer, and status lines show the student’s judgment before querying. Target delegation spends both answers on one emotion, while relation selection distributes them across emotions. The dialogue and uncertainty are illustrative. A dialogue links a missed train to feeling upset and an approaching train to relief. With two expert answers, querying both displayed candidates for the upset target leaves relief without expert help. Selecting uncertain relations instead distributes queries across both emotion targets.

Figure 2. Selective expert consultation over emotion–cause relations. The student scores every eligible target–candidate pair and queries near its F1 boundary fixed on validation. The expert receives the full dialogue and each selected relation. Fusion or direct replacement updates queried decisions while unqueried predictions remain unchanged. The dialogue, highlights, scores, and updates are illustrative, and every utterance remains a potential target. Indexed dialogue turns show candidate events and emotion expressions. The student scores every eligible emotion-target and candidate-cause relation, selects a bilateral boundary band, and chooses KEEP or CONSULT. The expert receives the full dialogue and selected questions. The lower panel preserves unqueried decisions, integrates queried judgments, and displays cause-to-emotion links.

(a) Candidate coverage
Dataset	Candidates $N$	$G_{\mathrm{in}}/G$	Coverage (%)
ECF	16,016	1,843/1,873	98.40
MEC 4	40,462	2,084/2,112	98.67

Table 1. Test coverage and nominal 10% queries.

G_{\mathrm{in}}/G

is gold inside the window over all gold. Across seeds 42–44,

p_{s}^{*}=\sigma(\tau_{s})

gives the threshold range,

Q

the mean answers, and

Q^{+}

the mean percentage of queries on the accepted student side.

Dataset	Model	Actual query %	Student ( $\approx$ 4B)	Random F	Selective F	Random R	Selective R	Expert ( $\geq$ 27B)
ECF	Qwen	9.31	$62.05_{\scriptscriptstyle\pm 0.77}$	$62.41_{\scriptscriptstyle\pm 0.79}$	$64.89_{\scriptscriptstyle\pm 0.79}$	$62.52_{\scriptscriptstyle\pm 0.66}$	$65.40_{\scriptscriptstyle\pm 0.64}$	$\mathbf{66.60}$
ECF	Gemma	9.08	$61.26_{\scriptscriptstyle\pm 0.66}$	$61.79_{\scriptscriptstyle\pm 0.59}$	$65.14_{\scriptscriptstyle\pm 0.29}$	$61.86_{\scriptscriptstyle\pm 0.56}$	$65.21_{\scriptscriptstyle\pm 0.05}$	$\mathbf{66.70}$
MEC 4	Qwen	9.65	$47.10_{\scriptscriptstyle\pm 0.64}$	$47.44_{\scriptscriptstyle\pm 0.63}$	$49.74_{\scriptscriptstyle\pm 0.60}$	$47.46_{\scriptscriptstyle\pm 0.61}$	$\mathbf{49.90}_{\scriptscriptstyle\pm 0.31}$	49.79
MEC 4	Gemma	9.65	$45.50_{\scriptscriptstyle\pm 0.31}$	$45.92_{\scriptscriptstyle\pm 0.16}$	$49.02_{\scriptscriptstyle\pm 0.86}$	$45.95_{\scriptscriptstyle\pm 0.25}$	$49.60_{\scriptscriptstyle\pm 0.36}$	$\mathbf{50.27}$

Table 2. Test pair F1, mean ±SD over three SFT seeds. Selective F (fusion) and Selective R (replacement) both use nominal 10% access. Random averages 20 draws with query counts matched in each dialogue. Row maxima are bold.

Method	Input	ECF	MEC 4
MECPE-2steps ( Wang et al., 2023 )	T	52.71	25.31
HiLo ( Li et al., 2025 )	T	53.09	34.22
M3F ( Wu et al., 2025 )	T	56.52	43.21
SCALE ( Ma et al., 2026 )	T	57.70	–
BridgeProt-SFT ( Xia et al., 2026a )	T	53.51	40.53
Jud-Local ( Feng et al., 2026 )	T	62.14	–

Table 3. Pair F1 under different backbones and source protocols. T/A/V denote text/audio/video. MECPE-2steps and HiLo follow ( Wu et al., 2025 ) , and other baselines follow their original reports. Consultation uses nominal 10% access (three-seed mean ±SD ). Dataset maxima are bold, and dashes denote unreported results.

(a) Query ranking: pair F1 and paired difference
Dataset	Model	$H_{\ell}$	$H_{p}$	Learned	Boundary	$\Delta$ vs. Learned [95% CI]
ECF	Qwen	$64.09_{\scriptscriptstyle\pm 0.59}$	$63.58_{\scriptscriptstyle\pm 0.60}$	$64.47_{\scriptscriptstyle\pm 0.50}$	$\mathbf{64.89}_{\scriptscriptstyle\pm 0.79}$	$+0.42\ [+0.04,+0.77]$
ECF	Gemma	$64.69_{\scriptscriptstyle\pm 0.67}$	$64.18_{\scriptscriptstyle\pm 0.38}$	$64.68_{\scriptscriptstyle\pm 0.50}$	$\mathbf{65.14}_{\scriptscriptstyle\pm 0.29}$	$+0.47\ [+0.14,+0.81]$
MEC 4	Qwen	$49.40_{\scriptscriptstyle\pm 0.96}$	$49.41_{\scriptscriptstyle\pm 0.96}$	$49.47_{\scriptscriptstyle\pm 0.45}$	$\mathbf{49.74}_{\scriptscriptstyle\pm 0.60}$	$+0.27\ [-0.02,+0.56]$
MEC 4	Gemma	$48.64_{\scriptscriptstyle\pm 0.92}$	$48.59_{\scriptscriptstyle\pm 0.82}$	$48.56_{\scriptscriptstyle\pm 0.79}$	$\mathbf{49.02}_{\scriptscriptstyle\pm 0.86}$	$+0.46\ [+0.21,+0.72]$

Table 4. Query allocation at nominal 10% access with fixed fusion and dialogue counts matched per comparison. (a) Test F1, mean ±SD , and Boundary minus Learned.

H_{\ell},H_{p}

denote native and calibrated entropy. Best F1 is bold. (b) Bilateral minus each restricted selector. Negative/positive denote rejected/accepted student predictions. Differences include pointwise 95% paired intervals from dialogue resampling.

(a) Pair F1
Dataset	Model	Relation	Target	Dialogue
ECF	Qwen	$\mathbf{64.95}_{\scriptscriptstyle\pm 0.81}$	$63.08_{\scriptscriptstyle\pm 0.63}$	$62.70_{\scriptscriptstyle\pm 1.02}$
ECF	Gemma	$\mathbf{65.22}_{\scriptscriptstyle\pm 0.47}$	$62.71_{\scriptscriptstyle\pm 0.50}$	$61.86_{\scriptscriptstyle\pm 0.93}$
MEC 4	Qwen	$\mathbf{49.77}_{\scriptscriptstyle\pm 0.59}$	$47.97_{\scriptscriptstyle\pm 0.60}$	$47.51_{\scriptscriptstyle\pm 0.60}$
MEC 4	Gemma	$\mathbf{49.06}_{\scriptscriptstyle\pm 0.87}$	$46.55_{\scriptscriptstyle\pm 0.44}$	$45.74_{\scriptscriptstyle\pm 0.28}$
(b) Complete multi-cause sets (%)

Table 5. Delegation under a hard 10% answer cap with fixed fusion, mean ±SD across three seeds. (a) Test pair F1. (b) Exact cause-set recovery (%) for targets with at least two gold causes, with 417 on ECF and 218 on MEC 4 . Row maxima are bold.

Figure 3. Expert access and test pair F1 across three SFT seeds. Points show means. Inset rows report mean ±SD at nominal 0/5/10/20% access. The horizontal references distinguish full fusion at the student threshold from the independent expert. Curves use actual query fractions. Random fusion matches query counts within every dialogue. Four panels plot mean F1 for selective fusion, direct replacement, and random fusion. Inset tables report means and sample standard deviations across three student seeds, including the zero-query endpoint. Separate horizontal lines mark full fusion and the expert. A single legend is above the panels.

Model	Student	Expert	Full fusion	Selective
Qwen	$15.57_{\scriptscriptstyle\pm 0.10}$	$61.58_{\scriptscriptstyle\pm 0.16}$	$76.88_{\scriptscriptstyle\pm 0.07}$	$27.60_{\scriptscriptstyle\pm 0.02}$
Gemma	$70.49_{\scriptscriptstyle\pm 0.14}$	$338.33_{\scriptscriptstyle\pm 0.38}$	$461.67_{\scriptscriptstyle\pm 1.66}$	$154.27_{\scriptscriptstyle\pm 0.11}$

Table 6. Online time in seconds, mean ±SD over three warmed repeats. Selective fusion uses nominal 10% access. All methods process the same 1,847 relations.

Dataset	Model	0–5%	5–10%	10–20%	20–100%
ECF	Qwen	$\mathbf{10.43}$	$6.92$	$1.32$	$0.01$
ECF	Gemma	$\mathbf{16.25}$	$6.40$	$1.87$	$0.02$
MEC 4	Qwen	$\mathbf{5.02}$	$2.96$	$0.11$	$0.00$
MEC 4	Gemma	$\mathbf{8.77}$	$1.89$	$0.47$	$0.00$

Table 7. Net correct revisions per 100 queries under fixed fusion, pooled across three seeds. Bins use validation distance quantiles outward from the student threshold. Row maxima are bold.

(a) Two ordinary SFT passes (seed 42)
Dataset	Model	Student	Random	Selective
ECF	Qwen	62.66	63.06	$\mathbf{65.89}$
ECF	Gemma	62.59	63.00	$\mathbf{65.77}$
MEC 4	Qwen	47.73	47.95	$\mathbf{49.70}$
MEC 4	Gemma	46.88	47.30	$\mathbf{50.16}$
(b) One SFT pass plus one KD pass (three seeds)

Table 8. Consultation after further training with the original expert. Fixed fusion uses nominal 10% access in (a,b) and a global 10% cap in (c). Values are test F1, with mean ±SD in (b).

H_{\ell},H_{p}

denote native and calibrated entropy. Row maxima are bold.

Appendix figures & tables39 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Split	D	U	N	G	O
ECF	Train	1,001	9,966	62,604	7,055	122
ECF	Validation	112	1,087	6,760	866	17
ECF	Test	261	2,566	16,016	1,873	30
MEC 4	Train	1,277	17,513	117,158	5,777	62
MEC 4	Validation	404	5,541	37,085	1,847	23
MEC 4	Test	443	6,051	40,462	2,112	28

Appendix

Table S1. Full dataset statistics for the P5/F2 candidate window. D, U, N, and G denote dialogues, utterances, candidates, and gold pairs. O counts gold pairs outside the window, retained as false negatives.

Dataset	Model	Seed	Val F1	Test F1	Test P	Test R
ECF	Qwen	42	67.42	66.60	62.06	71.86
ECF	Gemma	42	67.45	66.70	64.80	68.71
MEC 4	Qwen	42	48.74	49.79	42.37	60.37
MEC 4	Gemma	42	48.80	50.27	47.98	52.79

Appendix

Table S2. Fixed expert models. All use seed 42 and one SFT data pass. Each checkpoint is shared across the three student seeds in its setting.

Endpoint	yes	no	pad	Revision
Qwen 4B	9405	2083	248044	851bf6e806ef
Qwen 27B	9405	2083	248044	fc05daec18b0
Gemma E4B	4443	1904	0	ee0ef6023621
Gemma 31B	4443	1904	0	842da3794eaa

Appendix

Table S3. Answer token identities from the exact local tokenizer revisions. Full hashes and serialized examples are supplied in the evidence files.

Dataset	Model	Seed	$\tau_{s}$	$\tau_{e}$	$\tau_{f}$	$\epsilon_{.05}$	$\epsilon_{.10}$	$\epsilon_{.20}$
ECF	Qwen	42	-0.50	-0.80	-0.70	0.2943	0.5988	1.2278
ECF	Qwen	43	-0.95	-0.80	-0.75	0.2553	0.5546	1.0652
ECF	Qwen	44	-1.00	-0.80	-0.85	0.2417	0.5689	1.1142
ECF	Gemma	42	-0.75	-0.70	-0.60	0.3254	0.5372	1.1881
ECF	Gemma	43	-0.60	-0.70	-0.70	0.2121	0.5166	1.2271
ECF	Gemma	44	-0.55	-0.70	-0.70	0.3025	0.6413	1.2059

Appendix

Table S4. Frozen logit thresholds and query bands.

\tau_{f}

is selected separately on validation for full fusion. The fixed-fusion access curves instead retain

\tau_{s}

at every query rate, including full access. Bands are rounded here, with full precision supplied in JSON.

Dataset	Model	Seed	$p_{s}^{*}$	$Q$	Accepted	Rejected	Accepted (%)
ECF	Qwen	42	0.3775	1,544	675	869	43.72
ECF	Qwen	43	0.2789	1,447	662	785	45.75
ECF	Qwen	44	0.2689	1,480	689	791	46.55
ECF	Gemma	42	0.3208	1,428	661	767	46.29
ECF	Gemma	43	0.3543	1,395	709	686	50.82
ECF	Gemma	44	0.3659	1,539	600	939	38.99

Appendix

Table S5. Test query composition at nominal 10% access.

p_{s}^{*}=\sigma(\tau_{s})

is the calibrated decision probability. Accepted and rejected refer to the student’s predictions before consultation. Their counts sum to the actual answer count

Q

. Accepted (%) uses

Q

as its denominator.

Dataset	Model	Seed	Val student	Val selective F	Test student	Test random F	Test selective F	Test selective R
ECF	Qwen	42	63.05	65.32	61.28	61.60	64.00	64.75
ECF	Qwen	43	63.12	65.31	62.04	62.46	65.11	65.41
ECF	Qwen	44	63.82	65.52	62.82	63.18	65.55	66.03
ECF	Gemma	42	61.78	64.52	61.35	61.84	64.88	65.18
ECF	Gemma	43	62.25	66.08	61.87	62.35	65.46	65.27
ECF	Gemma	44	61.63	65.76	60.56	61.16	65.09	65.16

Appendix

Table S6. Primary nominal 10% results by independent student seed. Numbers are pair F1. Random results average 20 draws with query counts matched in each dialogue.

Dataset	Model	Seed	$\Delta$ student	95% interval	$\Delta$ random F	95% interval
ECF	Qwen	42	+2.72	[1.61, 3.84]	+2.41	[1.40, 3.40]
ECF	Qwen	43	+3.06	[2.14, 3.97]	+2.65	[1.85, 3.46]
ECF	Qwen	44	+2.73	[1.84, 3.57]	+2.37	[1.58, 3.08]
ECF	Gemma	42	+3.53	[2.59, 4.45]	+3.04	[2.19, 3.87]
ECF	Gemma	43	+3.59	[2.64, 4.59]	+3.11	[2.25, 4.00]
ECF	Gemma	44	+4.53	[3.33, 5.75]	+3.93	[2.89, 4.99]

Appendix

Table S7. Test differences at nominal 10% fusion. All intervals are conditional 95% paired bootstrap intervals over dialogues. Positive values favor selective fusion.

Dataset	Model	System	Emotion	Cause
ECF	Qwen	Student	$79.07_{\scriptscriptstyle\pm 0.60}$	$75.52_{\scriptscriptstyle\pm 0.71}$
		Fusion	$80.72_{\scriptscriptstyle\pm 0.67}$	$76.91_{\scriptscriptstyle\pm 0.32}$
		Replacement	$\mathbf{80.99}_{\scriptscriptstyle\pm 0.34}$	$\mathbf{77.43}_{\scriptscriptstyle\pm 0.23}$
ECF	Gemma	Student	$78.17_{\scriptscriptstyle\pm 0.71}$	$74.69_{\scriptscriptstyle\pm 0.37}$
		Fusion	$80.14_{\scriptscriptstyle\pm 0.65}$	$77.41_{\scriptscriptstyle\pm 0.40}$
		Replacement	$\mathbf{80.32}_{\scriptscriptstyle\pm 0.48}$	$\mathbf{77.41}_{\scriptscriptstyle\pm 0.38}$

Appendix

Table S8. Emotion and cause endpoint extraction at nominal 10% access, mean ±SD test F1 across three SFT seeds. Distinct endpoints are projected from pair outputs and full gold pairs. Bold marks the best endpoint result in each setting.

			Emotion endpoint F1			Cause endpoint F1
Dataset	Model	Seed	Student	F	R	Student	F	R
ECF	Qwen	42	78.39	79.96	80.68	74.98	76.64	77.20
ECF	Qwen	43	79.30	81.01	80.96	75.25	76.83	77.44
ECF	Qwen	44	79.53	81.19	81.35	76.33	77.26	77.65
ECF	Gemma	42	78.92	80.88	80.86	74.49	77.00	77.18
ECF	Gemma	43	78.06	79.71	80.15	75.11	77.79	77.85

Appendix

Table S9. Test endpoint F1 for each student. F and R denote nominal 10% selective fusion and replacement. Gold and predicted endpoints are projected from pair sets.

Dataset	Model	Slice	Gold	Student	Selective F	Selective R
ECF	Qwen	self	957	$71.46_{\scriptscriptstyle\pm 0.71}$	$73.49_{\scriptscriptstyle\pm 0.68}$	$74.03_{\scriptscriptstyle\pm 0.26}$
ECF	Qwen	external	916	$50.46_{\scriptscriptstyle\pm 3.23}$	$54.03_{\scriptscriptstyle\pm 3.45}$	$54.55_{\scriptscriptstyle\pm 2.02}$
ECF	Qwen	previous turn	534	$60.10_{\scriptscriptstyle\pm 1.55}$	$62.73_{\scriptscriptstyle\pm 1.62}$	$62.91_{\scriptscriptstyle\pm 0.80}$
ECF	Qwen	earlier by at least two	302	$32.02_{\scriptscriptstyle\pm 8.65}$	$38.98_{\scriptscriptstyle\pm 7.72}$	$40.29_{\scriptscriptstyle\pm 5.51}$
ECF	Qwen	future	80	$24.88_{\scriptscriptstyle\pm 7.68}$	$28.83_{\scriptscriptstyle\pm 6.25}$	$29.25_{\scriptscriptstyle\pm 2.71}$
ECF	Gemma	self	957	$70.56_{\scriptscriptstyle\pm 0.84}$	$74.13_{\scriptscriptstyle\pm 0.50}$	$74.55_{\scriptscriptstyle\pm 0.05}$

Appendix

Table S10. Test distance slices at nominal 10%, mean ±SD over three students. Gold counts refer to the dataset, not the sum across seeds.

Dataset	Model	$b$ (%)	Group	$Q$	Group F1	Relation F1	$\Delta$ [95% CI]
ECF	Qwen	5	Target	800.0	$62.57_{\scriptscriptstyle\pm 0.54}$	$63.98_{\scriptscriptstyle\pm 0.71}$	+1.41 [+1.00, +1.79]
ECF	Qwen	5	Dialogue	799.3	$62.32_{\scriptscriptstyle\pm 0.85}$	$63.98_{\scriptscriptstyle\pm 0.71}$	+1.66 [+1.12, +2.16]
ECF	Qwen	10	Target	1601.0	$63.08_{\scriptscriptstyle\pm 0.63}$	$64.95_{\scriptscriptstyle\pm 0.81}$	+1.88 [+1.37, +2.36]
ECF	Qwen	10	Dialogue	1599.7	$62.70_{\scriptscriptstyle\pm 1.02}$	$64.95_{\scriptscriptstyle\pm 0.82}$	+2.25 [+1.59, +2.89]
ECF	Qwen	20	Target	3203.0	$63.61_{\scriptscriptstyle\pm 0.92}$	$65.51_{\scriptscriptstyle\pm 0.80}$	+1.90 [+1.36, +2.42]
ECF	Qwen	20	Dialogue	3200.3	$63.28_{\scriptscriptstyle\pm 1.05}$	$65.51_{\scriptscriptstyle\pm 0.80}$	+2.24 [+1.58, +2.90]

Appendix

Table S11. Actual group access and the relation comparator at exactly that access, under fixed fusion.

\Delta

: matched Relation minus Group.

Dataset	Model	Gold target subset	$n$	Relation	Target	Dialogue
ECF	Qwen	At least one cause	1344	$\mathbf{48.26}_{\scriptscriptstyle\pm 0.55}$	$45.86_{\scriptscriptstyle\pm 0.71}$	$45.56_{\scriptscriptstyle\pm 0.62}$
ECF	Qwen	At least two causes	417	$\mathbf{38.85}_{\scriptscriptstyle\pm 6.17}$	$35.73_{\scriptscriptstyle\pm 4.99}$	$35.09_{\scriptscriptstyle\pm 5.48}$
ECF	Gemma	At least one cause	1344	$\mathbf{48.29}_{\scriptscriptstyle\pm 0.65}$	$45.16_{\scriptscriptstyle\pm 0.76}$	$44.64_{\scriptscriptstyle\pm 0.84}$
ECF	Gemma	At least two causes	417	$\mathbf{34.69}_{\scriptscriptstyle\pm 2.45}$	$31.65_{\scriptscriptstyle\pm 2.59}$	$31.25_{\scriptscriptstyle\pm 2.56}$
MEC 4	Qwen	At least one cause	1849	$\mathbf{37.79}_{\scriptscriptstyle\pm 0.19}$	$34.36_{\scriptscriptstyle\pm 1.68}$	$33.80_{\scriptscriptstyle\pm 2.08}$
MEC 4	Qwen	At least two causes	218	$\mathbf{18.81}_{\scriptscriptstyle\pm 2.75}$	$17.28_{\scriptscriptstyle\pm 1.74}$	$16.67_{\scriptscriptstyle\pm 2.26}$

Appendix

Table S12. Exact recovery of cause sets (%) at a hard 10% answer cap.

n

is the number of eligible gold targets, counted once per dataset. Row maxima are bold.

No.	Feature	Definition
1	Student score	Calibrated logit $z_{s}(i,j)$ .
2	Boundary distance	$\|z_{s}(i,j)-\tau_{s}\|$ .
3	Student class	$\mathbb{1}[z_{s}(i,j)>\tau_{s}]$ .
4	Signed distance	$j-i$ . Positive values denote later candidate utterances.
5	Speaker match	Equality of target and candidate speaker strings.
6	Candidate length	Character count of candidate utterance text.

Appendix

Table S13. The 16 features available to the learned revision value regressor.

\mathcal{C}_{i}

contains the target’s candidate relations and

n

is dialogue length.

Dataset	Model	Seed	F-M	F-G	R-M	R-G
ECF	Qwen	42	C/C/C	C/C/C	L/C/L	L/C/L
ECF	Qwen	43	C/C/C	C/C/C	C/C/C	C/C/L
ECF	Qwen	44	C/C/C	C/C/C	C/C/C	C/C/C
ECF	Gemma	42	C/C/L	C/C/C	C/C/C	C/C/C
ECF	Gemma	43	C/C/C	C/C/C	C/C/C	C/C/C
ECF	Gemma	44	C/C/C	C/C/C	C/C/C	C/C/C

Appendix

Table S14. Selected router targets at 5/10/20%. C denotes count improvement and L reduction in logistic loss. F and R denote fusion and replacement. M matches dialogue counts, and G uses a global hard budget.

Dataset	Model	$b$ (%)	$Q$	$H_{\ell}$	$H_{p}$	Learned	Boundary	$\Delta$ [95% CI]
ECF	Qwen	5	661.3	$62.81_{\scriptscriptstyle\pm 0.76}$	$62.44_{\scriptscriptstyle\pm 0.80}$	$63.63_{\scriptscriptstyle\pm 0.44}$	$\mathbf{63.69}_{\scriptscriptstyle\pm 0.63}$	+0.06 [-0.28, +0.40]
ECF	Qwen	10	1490.3	$64.09_{\scriptscriptstyle\pm 0.59}$	$63.58_{\scriptscriptstyle\pm 0.60}$	$64.47_{\scriptscriptstyle\pm 0.50}$	$\mathbf{64.89}_{\scriptscriptstyle\pm 0.79}$	+0.42 [+0.04, +0.77]
ECF	Qwen	20	3108.3	$65.36_{\scriptscriptstyle\pm 0.70}$	$65.32_{\scriptscriptstyle\pm 0.72}$	$64.96_{\scriptscriptstyle\pm 0.58}$	$\mathbf{65.51}_{\scriptscriptstyle\pm 0.80}$	+0.55 [+0.20, +0.89]
ECF	Gemma	5	714.0	$62.62_{\scriptscriptstyle\pm 1.32}$	$62.25_{\scriptscriptstyle\pm 0.51}$	$63.59_{\scriptscriptstyle\pm 0.28}$	$\mathbf{63.95}_{\scriptscriptstyle\pm 0.20}$	+0.36 [+0.00, +0.73]
ECF	Gemma	10	1454.0	$64.69_{\scriptscriptstyle\pm 0.67}$	$64.18_{\scriptscriptstyle\pm 0.38}$	$64.68_{\scriptscriptstyle\pm 0.50}$	$\mathbf{65.14}_{\scriptscriptstyle\pm 0.29}$	+0.47 [+0.14, +0.81]
ECF	Gemma	20	3094.0	$65.89_{\scriptscriptstyle\pm 0.45}$	$65.84_{\scriptscriptstyle\pm 0.40}$	$65.01_{\scriptscriptstyle\pm 0.71}$	$\mathbf{65.99}_{\scriptscriptstyle\pm 0.36}$	+0.98 [+0.57, +1.40]

Appendix

Table S15. Fusion with matched query counts per dialogue.

Q

: mean answers.

\Delta

: Boundary minus Learned with a 95% paired interval.

Dataset	Model	$b$ (%)	$Q$	$H_{\ell}$	$H_{p}$	Learned	Boundary	$\Delta$ [95% CI]
ECF	Qwen	5	800.0	$62.87_{\scriptscriptstyle\pm 0.62}$	$62.43_{\scriptscriptstyle\pm 0.61}$	$63.77_{\scriptscriptstyle\pm 0.49}$	$\mathbf{63.98}_{\scriptscriptstyle\pm 0.71}$	+0.21 [-0.17, +0.57]
ECF	Qwen	10	1601.0	$64.12_{\scriptscriptstyle\pm 0.69}$	$63.71_{\scriptscriptstyle\pm 0.64}$	$64.52_{\scriptscriptstyle\pm 0.62}$	$\mathbf{64.95}_{\scriptscriptstyle\pm 0.81}$	+0.44 [+0.04, +0.80]
ECF	Qwen	20	3203.0	$65.47_{\scriptscriptstyle\pm 0.80}$	$65.44_{\scriptscriptstyle\pm 0.81}$	$64.95_{\scriptscriptstyle\pm 0.55}$	$\mathbf{65.51}_{\scriptscriptstyle\pm 0.80}$	+0.56 [+0.20, +0.90]
ECF	Gemma	5	800.0	$62.69_{\scriptscriptstyle\pm 1.57}$	$61.95_{\scriptscriptstyle\pm 0.78}$	$63.78_{\scriptscriptstyle\pm 0.42}$	$\mathbf{64.06}_{\scriptscriptstyle\pm 0.19}$	+0.28 [-0.13, +0.72]
ECF	Gemma	10	1601.0	$64.77_{\scriptscriptstyle\pm 0.80}$	$64.32_{\scriptscriptstyle\pm 0.70}$	$64.91_{\scriptscriptstyle\pm 0.60}$	$\mathbf{65.22}_{\scriptscriptstyle\pm 0.47}$	+0.31 [-0.03, +0.67]
ECF	Gemma	20	3203.0	$65.93_{\scriptscriptstyle\pm 0.42}$	$65.92_{\scriptscriptstyle\pm 0.47}$	$65.20_{\scriptscriptstyle\pm 0.51}$	$\mathbf{65.99}_{\scriptscriptstyle\pm 0.35}$	+0.79 [+0.38, +1.22]

Appendix

Table S16. Fusion under a global hard budget.

Q

: mean answers.

\Delta

: Boundary minus Learned with a 95% paired interval.

Dataset	Model	$b$ (%)	$Q$	$H_{\ell}$	$H_{p}$	Learned	Boundary	$\Delta$ [95% CI]
ECF	Qwen	5	661.3	$62.77_{\scriptscriptstyle\pm 0.64}$	$62.46_{\scriptscriptstyle\pm 0.64}$	$63.70_{\scriptscriptstyle\pm 0.99}$	$\mathbf{63.92}_{\scriptscriptstyle\pm 0.58}$	+0.22 [-0.24, +0.66]
ECF	Qwen	10	1490.3	$64.47_{\scriptscriptstyle\pm 0.48}$	$64.00_{\scriptscriptstyle\pm 0.23}$	$64.69_{\scriptscriptstyle\pm 0.36}$	$\mathbf{65.40}_{\scriptscriptstyle\pm 0.64}$	+0.71 [+0.29, +1.11]
ECF	Qwen	20	3108.3	$66.30_{\scriptscriptstyle\pm 0.13}$	$66.21_{\scriptscriptstyle\pm 0.15}$	$65.12_{\scriptscriptstyle\pm 0.41}$	$\mathbf{66.33}_{\scriptscriptstyle\pm 0.05}$	+1.21 [+0.60, +1.77]
ECF	Gemma	5	714.0	$62.46_{\scriptscriptstyle\pm 1.44}$	$61.96_{\scriptscriptstyle\pm 0.48}$	$63.56_{\scriptscriptstyle\pm 0.48}$	$\mathbf{63.91}_{\scriptscriptstyle\pm 0.20}$	+0.35 [-0.08, +0.79]
ECF	Gemma	10	1454.0	$64.56_{\scriptscriptstyle\pm 0.62}$	$64.04_{\scriptscriptstyle\pm 0.17}$	$64.80_{\scriptscriptstyle\pm 0.59}$	$\mathbf{65.21}_{\scriptscriptstyle\pm 0.05}$	+0.40 [-0.13, +0.93]
ECF	Gemma	20	3094.0	$66.27_{\scriptscriptstyle\pm 0.32}$	$66.13_{\scriptscriptstyle\pm 0.23}$	$65.68_{\scriptscriptstyle\pm 0.45}$	$\mathbf{66.43}_{\scriptscriptstyle\pm 0.19}$	+0.75 [+0.26, +1.26]

Appendix

Table S17. Replacement with matched query counts per dialogue.

Q

: mean answers.

\Delta

: Boundary minus Learned with a 95% paired interval.

Dataset	Model	$b$ (%)	$Q$	$H_{\ell}$	$H_{p}$	Learned	Boundary	$\Delta$ [95% CI]
ECF	Qwen	5	800.0	$63.00_{\scriptscriptstyle\pm 0.69}$	$62.52_{\scriptscriptstyle\pm 0.58}$	$63.63_{\scriptscriptstyle\pm 1.12}$	$\mathbf{64.22}_{\scriptscriptstyle\pm 0.70}$	+0.59 [+0.08, +1.12]
ECF	Qwen	10	1601.0	$64.44_{\scriptscriptstyle\pm 0.68}$	$64.10_{\scriptscriptstyle\pm 0.55}$	$64.87_{\scriptscriptstyle\pm 0.39}$	$\mathbf{65.45}_{\scriptscriptstyle\pm 0.65}$	+0.58 [+0.14, +1.01]
ECF	Qwen	20	3203.0	$\mathbf{66.38}_{\scriptscriptstyle\pm 0.21}$	$66.24_{\scriptscriptstyle\pm 0.22}$	$65.21_{\scriptscriptstyle\pm 0.59}$	$66.32_{\scriptscriptstyle\pm 0.07}$	+1.11 [+0.48, +1.69]
ECF	Gemma	5	800.0	$62.58_{\scriptscriptstyle\pm 1.68}$	$61.66_{\scriptscriptstyle\pm 0.70}$	$63.84_{\scriptscriptstyle\pm 0.50}$	$\mathbf{64.01}_{\scriptscriptstyle\pm 0.16}$	+0.17 [-0.29, +0.67]
ECF	Gemma	10	1601.0	$64.70_{\scriptscriptstyle\pm 0.85}$	$64.20_{\scriptscriptstyle\pm 0.46}$	$64.98_{\scriptscriptstyle\pm 0.65}$	$\mathbf{65.28}_{\scriptscriptstyle\pm 0.39}$	+0.29 [-0.25, +0.84]
ECF	Gemma	20	3203.0	$66.37_{\scriptscriptstyle\pm 0.38}$	$66.35_{\scriptscriptstyle\pm 0.29}$	$65.63_{\scriptscriptstyle\pm 0.39}$	$\mathbf{66.44}_{\scriptscriptstyle\pm 0.21}$	+0.81 [+0.35, +1.32]

Appendix

Table S18. Replacement under a global hard budget.

Q

: mean answers.

\Delta

: Boundary minus Learned with a 95% paired interval.

Dataset	Model	Selector	$Q$	Restricted	Bilateral	$\Delta$ [95% CI]
ECF	Qwen	Negative only	1490.3	$63.68_{\scriptscriptstyle\pm 0.35}$	$64.89_{\scriptscriptstyle\pm 0.79}$	+1.21 [+0.76, +1.61]
ECF	Qwen	Positive only	1402.3	$63.79_{\scriptscriptstyle\pm 1.38}$	$64.88_{\scriptscriptstyle\pm 0.76}$	+1.09 [+0.54, +1.62]
ECF	Qwen	One rejected/target	1464.3	$63.61_{\scriptscriptstyle\pm 0.37}$	$64.85_{\scriptscriptstyle\pm 0.77}$	+1.24 [+0.79, +1.66]
ECF	Gemma	Negative only	1453.3	$64.03_{\scriptscriptstyle\pm 0.37}$	$65.14_{\scriptscriptstyle\pm 0.28}$	+1.10 [+0.50, +1.72]
ECF	Gemma	Positive only	1351.7	$62.96_{\scriptscriptstyle\pm 1.09}$	$64.99_{\scriptscriptstyle\pm 0.32}$	+2.03 [+1.36, +2.66]
ECF	Gemma	One rejected/target	1427.7	$63.41_{\scriptscriptstyle\pm 0.21}$	$65.10_{\scriptscriptstyle\pm 0.32}$	+1.69 [+1.07, +2.30]

Appendix

Table S19. Query side at nominal 10% access under fixed fusion. Both methods use the actual

Q

shown.

\Delta

is Bilateral minus the restricted selector.

Dataset	Model	Selector	$Q$	Restricted	Bilateral	$\Delta$ [95% CI]
ECF	Qwen	Negative only	661.3	$63.21_{\scriptscriptstyle\pm 0.43}$	$63.69_{\scriptscriptstyle\pm 0.63}$	+0.48 [+0.13, +0.82]
ECF	Qwen	Positive only	653.3	$63.41_{\scriptscriptstyle\pm 1.22}$	$63.69_{\scriptscriptstyle\pm 0.63}$	+0.28 [-0.15, +0.74]
ECF	Qwen	One rejected/target	660.0	$63.27_{\scriptscriptstyle\pm 0.44}$	$63.69_{\scriptscriptstyle\pm 0.62}$	+0.42 [+0.07, +0.75]
ECF	Gemma	Negative only	713.3	$63.23_{\scriptscriptstyle\pm 0.19}$	$63.95_{\scriptscriptstyle\pm 0.20}$	+0.71 [+0.26, +1.17]
ECF	Gemma	Positive only	701.3	$62.70_{\scriptscriptstyle\pm 1.05}$	$63.96_{\scriptscriptstyle\pm 0.21}$	+1.26 [+0.78, +1.72]
ECF	Gemma	One rejected/target	712.3	$63.05_{\scriptscriptstyle\pm 0.21}$	$63.95_{\scriptscriptstyle\pm 0.20}$	+0.90 [+0.43, +1.35]

Appendix

Table S20. Query side at nominal 5% access under fixed fusion. Both methods use the actual

Q

shown.

\Delta

is Bilateral minus the restricted selector.

Dataset	Model	Selector	$Q$	Restricted	Bilateral	$\Delta$ [95% CI]
ECF	Qwen	Negative only	3107.7	$63.76_{\scriptscriptstyle\pm 0.24}$	$65.51_{\scriptscriptstyle\pm 0.80}$	+1.75 [+1.29, +2.15]
ECF	Qwen	Positive only	2088.0	$63.82_{\scriptscriptstyle\pm 1.39}$	$65.16_{\scriptscriptstyle\pm 0.90}$	+1.34 [+0.73, +1.91]
ECF	Qwen	One rejected/target	2373.7	$63.61_{\scriptscriptstyle\pm 0.32}$	$65.24_{\scriptscriptstyle\pm 0.72}$	+1.63 [+1.20, +2.06]
ECF	Gemma	Negative only	3092.0	$64.20_{\scriptscriptstyle\pm 0.26}$	$65.98_{\scriptscriptstyle\pm 0.35}$	+1.78 [+1.19, +2.38]
ECF	Gemma	Positive only	1912.0	$62.99_{\scriptscriptstyle\pm 1.01}$	$65.35_{\scriptscriptstyle\pm 0.43}$	+2.36 [+1.68, +3.04]
ECF	Gemma	One rejected/target	2365.3	$63.45_{\scriptscriptstyle\pm 0.11}$	$65.83_{\scriptscriptstyle\pm 0.45}$	+2.38 [+1.75, +3.02]

Appendix

Table S21. Query side at nominal 20% access under fixed fusion. Both methods use the actual

Q

shown.

\Delta

is Bilateral minus the restricted selector.

Dataset	Model	Nominal query %	Random policy	Actual query %	Fusion	Replacement
ECF	Qwen	5	uniform matched total	4.13	$62.15_{\scriptscriptstyle\pm 0.77}$	$62.20_{\scriptscriptstyle\pm 0.72}$
ECF	Qwen	5	uniform nominal	5.00	$62.17_{\scriptscriptstyle\pm 0.76}$	$62.23_{\scriptscriptstyle\pm 0.71}$
ECF	Qwen	5	per-dialogue matched	4.13	$62.23_{\scriptscriptstyle\pm 0.74}$	$62.29_{\scriptscriptstyle\pm 0.69}$
ECF	Qwen	10	uniform matched total	9.31	$62.31_{\scriptscriptstyle\pm 0.76}$	$62.43_{\scriptscriptstyle\pm 0.67}$
ECF	Qwen	10	uniform nominal	10.00	$62.33_{\scriptscriptstyle\pm 0.77}$	$62.45_{\scriptscriptstyle\pm 0.66}$
ECF	Qwen	10	per-dialogue matched	9.31	$62.41_{\scriptscriptstyle\pm 0.79}$	$62.52_{\scriptscriptstyle\pm 0.66}$

Appendix

Table S22. Additional random query controls on test. Query fractions and F1 are group means, with

\pm

denoting sample SD across three students. All policies use the same 20 random query seeds.

Dataset	Model	Seed	$Q$	Boundary	Stratified	$\Delta$ [95% CI]
ECF	Qwen	42	1544	64.00	62.35	+1.65 [+0.88, +2.37]
ECF	Qwen	43	1447	65.11	63.12	+1.98 [+1.39, +2.57]
ECF	Qwen	44	1480	65.55	64.10	+1.44 [+0.87, +2.02]
ECF	Gemma	42	1428	64.88	62.93	+1.95 [+1.36, +2.56]
ECF	Gemma	43	1395	65.46	63.72	+1.74 [+1.09, +2.37]
ECF	Gemma	44	1539	65.09	62.72	+2.37 [+1.49, +3.22]

Appendix

Table S23. Stratified random control at nominal 10% access. Twenty random draws preserve counts jointly by dialogue, the student’s predicted class, and distance category (self, previous, earlier, or future).

\Delta

denotes Boundary minus Stratified.

Dataset	Model	F	Local F	R	Expert
ECF	Qwen	$65.51_{\scriptscriptstyle\pm 0.80}$	$66.31_{\scriptscriptstyle\pm 0.25}$	$66.33_{\scriptscriptstyle\pm 0.05}$	$\mathbf{66.60}$
ECF	Gemma	$65.99_{\scriptscriptstyle\pm 0.36}$	$66.58_{\scriptscriptstyle\pm 0.28}$	$66.43_{\scriptscriptstyle\pm 0.19}$	$\mathbf{66.70}$
MEC 4	Qwen	$49.90_{\scriptscriptstyle\pm 0.54}$	$\mathbf{50.28}_{\scriptscriptstyle\pm 0.08}$	$49.86_{\scriptscriptstyle\pm 0.09}$	$49.79$
MEC 4	Gemma	$49.45_{\scriptscriptstyle\pm 1.01}$	$\mathbf{50.28}_{\scriptscriptstyle\pm 0.25}$	$50.23_{\scriptscriptstyle\pm 0.04}$	$50.27$

Appendix

Table S24. Integration at nominal 20% access, mean ±SD test F1. F denotes fixed fusion, Local F fits weights and thresholds for queried relations, and R denotes replacement. Expert uses all relations. Row maxima are bold.

Dataset	Model	Seed	$\alpha$	$\tau_{q}$	F	Local F
ECF	Qwen	42	0.75	-0.65	64.59	66.02
ECF	Qwen	43	0.75	-0.75	65.87	66.50
ECF	Qwen	44	0.50	-0.85	66.05	66.40
ECF	Gemma	42	0.75	-0.65	65.69	66.28
ECF	Gemma	43	0.75	-0.55	66.38	66.61
ECF	Gemma	44	0.75	-0.70	65.89	66.84

Appendix

Table S25. Validation-selected Local F parameters and test F1 at nominal 20% access. F uses

\alpha=0.5

and

\tau_{s}

. Local F uses the listed parameters.

Policy	$q$ (%)	$D$	Time (s)	Tok. (k)	Pairs/s
Qwen
Student	0.00	0.0	$15.57_{\scriptscriptstyle\pm 0.10}$	181.5	118.7
Expert	100.00	32.0	$61.58_{\scriptscriptstyle\pm 0.16}$	168.5	30.0
Full fusion	100.00	32.0	$76.88_{\scriptscriptstyle\pm 0.07}$	350.0	24.0
Sel. 5%	3.95	25.0	$23.71_{\scriptscriptstyle\pm 0.01}$	195.6	77.9
Sel. 10%	9.42	30.0	$27.60_{\scriptscriptstyle\pm 0.02}$	206.6	66.9

Appendix

Table S26. Online execution on 1,847 relations. Time is mean ±SD over three warmed repeats.

D

counts dialogues with at least one expert query out of 32. Tokens count processed input positions, including padding and repeated prefixes. k denotes thousands.

Model	Policy	Allocated	Reserved	Diff
Qwen	Student	10.66	16.61	0
Qwen	Expert	52.05	53.36	0
Qwen	Full fusion	60.83	67.48	0
Qwen	Selective 5%	60.83	67.48	2
Qwen	Selective 10%	60.83	67.48	3
Qwen	Selective 20%	60.83	67.48	2

Appendix

Table S27. Peak PyTorch allocated and reserved GPU memory in GiB over the three formal repeats. Diff counts final predictions differing from cached-score replay at the same queries. No unqueried decision changes relative to the online student.

Dataset	Model	Bin (%)	$Q$	F repair	F damage	R repair	R damage
ECF	Qwen	0–5	661.3	23.59	13.16	26.61	15.02
ECF	Qwen	5–10	829.0	12.83	5.91	20.59	11.06
ECF	Qwen	10–20	1618.0	2.58	1.26	8.47	6.26
ECF	Qwen	20–100	12907.7	0.02	0.01	0.47	0.50
ECF	Gemma	0–5	714.0	27.78	11.53	30.44	14.19
ECF	Gemma	5–10	740.0	12.57	6.17	19.41	12.25

Appendix

Table S28. Repairs and damage per 100 queries by validation distance bin. Counts are pooled across three seeds and

Q

is mean bin size. F denotes fixed fusion and R replacement.

Table 40

Dataset	Model	$b$	$Q$	Student	Random F	Selective F	Selective R	$\Delta$ [95% CI]
ECF	Qwen	5	723	62.66	62.89	64.78	64.69	+1.89 [+1.21, +2.58]
ECF	Qwen	10	1,458	62.66	63.06	65.89	65.77	+2.83 [+2.00, +3.66]
ECF	Qwen	20	3,211	62.66	63.51	66.40	66.73	+2.89 [+2.10, +3.67]
ECF	Gemma	5	765	62.59	62.80	64.54	64.39	+1.74 [+1.01, +2.47]
ECF	Gemma	10	1,516	62.59	63.00	65.77	65.71	+2.77 [+1.97, +3.63]
ECF	Gemma	20	3,089	62.59	63.34	66.45	66.74	+3.12 [+2.31, +3.93]

Appendix

Table S31. Consultation after two ordinary SFT passes, seed 42, with the fixed expert from the first pass.

b

is nominal query percentage and

Q

actual answers. Scores are test pair F1.

\Delta

compares selective fusion with the mean of 20 random draws matched by dialogue counts. Intervals use 2,000 paired dialogue resamples.

(a) Test
Dataset	Model	$Q$	Student	$H_{\ell}$	$H_{p}$	Learned	Boundary
ECF	Qwen	1,601	62.66	64.40	64.15	65.16	$\mathbf{65.98}$
ECF	Gemma	1,601	62.59	65.00	65.02	64.81	$\mathbf{65.74}$
MEC 4	Qwen	4,046	47.73	49.49	49.39	49.15	$\mathbf{49.69}$
MEC 4	Gemma	4,046	46.88	50.08	50.06	49.30	$\mathbf{50.15}$
(b) Validation

Appendix

Table S32. Query rankings after two ordinary SFT passes, seed 42, under a hard global 10% cap with fixed fusion.

Q

is the actual number of expert answers per policy.

H_{\ell}

and

H_{p}

denote native and calibrated entropy. Validation uses out-of-fold predictions for Learned. Scores are pair F1, with row maxima in bold.

Dataset	Model	Boundary $-$ $H_{\ell}$	Boundary $-$ $H_{p}$	Boundary $-$ Learned
ECF	Qwen	+1.58 [+0.84, +2.35]	+1.83 [+1.07, +2.61]	+0.82 [+0.31, +1.35]
ECF	Gemma	+0.74 [+0.21, +1.29]	+0.72 [+0.18, +1.25]	+0.93 [+0.35, +1.52]
MEC 4	Qwen	+0.20 [-0.20, +0.61]	+0.30 [-0.14, +0.75]	+0.54 [+0.04, +1.06]
MEC 4	Gemma	+0.07 [-0.29, +0.46]	+0.09 [-0.31, +0.49]	+0.85 [+0.40, +1.32]

Appendix

Table S33. Boundary minus each competing selector after two ordinary SFT passes, seed 42. All use fixed fusion and the same hard global 10% budget. Entries are test F1 differences with pointwise 95% paired intervals from 2,000 dialogue resamples, conditional on the trained models and frozen calibration.

Dataset	Model	Seed	Val student	Test student	Random F	Selective F	$\Delta$ [95% CI]
ECF	Qwen	42	66.06	65.04	65.17	66.73	+1.56 [+0.76, +2.37]
ECF	Qwen	43	65.45	64.86	65.02	66.23	+1.21 [+0.38, +2.03]
ECF	Qwen	44	65.91	64.95	65.14	66.49	+1.35 [+0.55, +2.18]
ECF	Gemma	42	64.40	63.25	63.52	64.75	+1.23 [+0.37, +2.17]
ECF	Gemma	43	64.76	64.01	64.30	66.34	+2.04 [+1.18, +2.93]
ECF	Gemma	44	64.25	64.12	64.60	66.97	+2.37 [+1.55, +3.20]

Appendix

Table S34. Additional KD condition at nominal 10% access. All students have one SFT pass plus one KD pass.

\Delta

: selective fusion minus matched random fusion.

Dataset	Student	$b$ (%)	Same F	Cross F	Cross R	$\Delta$ [95% CI]
ECF	Qwen	5	$63.69_{\scriptscriptstyle\pm 0.63}$	$64.04_{\scriptscriptstyle\pm 0.30}$	$64.19_{\scriptscriptstyle\pm 0.35}$	+0.35 [+0.02, +0.70]
ECF	Qwen	10	$64.89_{\scriptscriptstyle\pm 0.79}$	$65.45_{\scriptscriptstyle\pm 0.34}$	$65.54_{\scriptscriptstyle\pm 0.53}$	+0.56 [+0.02, +1.14]
ECF	Qwen	20	$65.51_{\scriptscriptstyle\pm 0.80}$	$66.16_{\scriptscriptstyle\pm 0.35}$	$66.69_{\scriptscriptstyle\pm 0.14}$	+0.66 [+0.03, +1.32]
ECF	Gemma	5	$63.95_{\scriptscriptstyle\pm 0.20}$	$63.49_{\scriptscriptstyle\pm 0.37}$	$63.86_{\scriptscriptstyle\pm 0.12}$	-0.46 [-0.86, -0.07]
ECF	Gemma	10	$65.14_{\scriptscriptstyle\pm 0.29}$	$64.53_{\scriptscriptstyle\pm 0.44}$	$65.01_{\scriptscriptstyle\pm 0.21}$	-0.61 [-1.18, -0.05]
ECF	Gemma	20	$65.99_{\scriptscriptstyle\pm 0.36}$	$65.17_{\scriptscriptstyle\pm 0.45}$	$66.33_{\scriptscriptstyle\pm 0.24}$	-0.82 [-1.54, -0.07]

Appendix

Table S35. Experts from another family, mean ±SD test F1. Student identifies the querying model. Cross uses the other family’s expert with student queries unchanged.

\Delta

is Cross F minus Same F.

Dataset	Model	Seed	Mean $\Delta$	2.5–97.5%	Minimum	Positive (%)	Query range (%)
ECF	Qwen	42	+2.62	[+2.20, +3.27]	+1.95	100	[8.62, 10.72]
ECF	Qwen	43	+3.01	[+2.37, +3.58]	+1.86	100	[8.24, 10.21]
ECF	Qwen	44	+2.82	[+2.12, +3.74]	+2.07	100	[8.59, 10.48]
ECF	Gemma	42	+3.87	[+3.19, +4.38]	+2.86	100	[8.19, 10.38]
ECF	Gemma	43	+3.67	[+2.89, +4.48]	+2.59	100	[8.14, 10.22]
ECF	Gemma	44	+3.64	[+2.29, +4.53]	+1.70	100	[8.44, 10.73]

Appendix

Table S36. Calibration sensitivity at nominal 10% access. Columns report consultation gain over the recalibrated student, its central range and minimum, the fraction of positive draws, and the actual access range.

Table 47

Table 48

Table 49

Table 50

Evidence Selection

Evidence selection chooses the documents, passages, claims, or observations that best support an answer or decision.

See all 60 papers

Jul 29, 2026

Generation or Judgement? A Paradigm Perspective on LLM-Based Emotion-Cause Pair Extraction in Conversation

Emotion-cause pair extraction in conversation (ECPEC) identifies utterance pairs in which one utterance causes an emotion expressed in another. Recent LLM-based approaches formulate ECPEC at markedly different granularities, ranging from generating complete pair sets to judging individual candidate pairs. In this paper, we make the surprising observation that task formulation substantially affects performance, where pair-level judgement outperforms dialogue-level generation in all 18 controlled comparisons. We investigate the sources of this paradigm gap and find that many relations omitted by dialogue-level generation remain recognizable under explicit pair queries, under which the model recognizes 92.7%-98.1% of emotion-cause relations. This suggests that LLMs can recognize emotion-cause relations but struggle to discover and return complete pair sets. Pair-level judgement alleviates this burden, although its candidate rankings are more reliable than the binary decisions produced by a shared threshold. Based on this diagnosis, we introduce an auxiliary retriever that selectively re-examines ambiguous boundary cases, yielding consistent F1 improvements of 0.50-1.46 points across three datasets while maintaining an inference time of only 1.49x that of the baseline paradigm. These findings show that task decomposition and candidate scope are critical to effectively utilizing LLMs for ECPEC.

Weijie Feng, Hongchuang Wang, Binbin Liu +1

Hefei University of Technology

Relation Extraction Emotion Recognition in Conversations

Ask PDF

Apr 21, 2026

Emotion-Cause Pair Extraction in Conversations via Semantic Decoupling and Graph Alignment

Emotion-Cause Pair Extraction in Conversations (ECPEC) aims to identify the set of causal relations between emotion utterances and their triggering causes within a dialogue. Most existing approaches formulate ECPEC as an independent pairwise classification task, overlooking the distinct semantics of emotion diffusion and cause explanation, and failing to capture globally consistent many-to-many conversational causality. To address these limitations, we revisit ECPEC from a semantic perspective and seek to disentangle emotion-oriented semantics from cause-oriented semantics, mapping them into two complementary representation spaces to better capture their distinct conversational roles. Building on this semantic decoupling, we naturally formulate ECPEC as a global alignment problem between the emotion-side and cause-side representations, and employ optimal transport to enable many-to-many and globally consistent emotion-cause matching. Based on this perspective, we propose a unified framework SCALE that instantiates the above semantic decoupling and alignment principle within a shared conversational structure. Extensive experiments on several benchmark datasets demonstrate that SCALE consistently achieves state-of-the-art performance. Our codes are released at https://github.com/CoCoSphere/SCALE.

Tianxiang Ma, Weijie Feng, Xinyu Wang +1

Hefei University of Technology

Relation Extraction Emotion Recognition in Conversations

Ask PDF

Jun 17, 2026

Learning Robust Pair Confidence for Multimodal Emotion-Cause Pair Extraction

Multimodal emotion-cause pair extraction (MECPE) requires reliable pair confidence over candidate pairs. Existing pair scorers commonly use pair-level cross entropy over valid candidates, which treats links mostly independently. This leaves the relative confidence geometry among competing causes under-constrained, allowing gold pairs to stay close to hard negatives or rely on incidental non-gold context. We study this vulnerability as pair-confidence brittleness and propose RPCL (Robust Pair Confidence Learning), a training-only framework for pair-confidence learning. RPCL encourages pair confidence to be both discriminative and stable: gold pairs are separated from row-wise hard negatives through a confidence-difference margin constraint, and clean pair predictions are aligned with predictions from a corrupted view where non-gold contextual utterance representations are partially corrupted. The original clean pair scorer and decoding pipeline are used unchanged at inference time. On ECF, MECAD, and MEC4, RPCL improves the three-seed mean Pair F1 over a matched base model by 2.58 to 2.83 percentage points in the full text-audio-video setting, and improves mean Pair AUPRC on all three datasets. Diagnostic analysis further shows larger gold-negative confidence gaps and lower margin-violation severity. These results suggest that explicitly shaping pair confidence is an effective training strategy for MECPE.

Zhuangzhuang Pan, Ning Dong, Yingna Su +1

Institute for Advanced Studies Universiti Malaya Kuala Lumpur 50603, Malaysia · School of Information Engineering Suqian University Suqian 223800, China · Digitization Department Suzhou University of Technology Suzhou 215500, China

Pairwise Preference Learning Multimodal Robustness Emotion Recognition in Conversations

Ask PDF

Abstract

Evidence Selection

Explore similar work

Generation or Judgement? A Paradigm Perspective on LLM-Based Emotion-Cause Pair Extraction in Conversation

Emotion-Cause Pair Extraction in Conversations via Semantic Decoupling and Graph Alignment

Learning Robust Pair Confidence for Multimodal Emotion-Cause Pair Extraction