PaperlayerView exact revision on arXiv

Source-linked AI summary

Improving JEV Entailment Reranking with Span Readouts and Selective Adaptation

Haocheng Mei, Yuchuan Tian, Ye Yuan, Enhua Wu, Chao Xu, Allen Chen, Kai Han, Yunhe Wang

arXiv:2610.35654v1cs.CL

TL;DR

The paper asks whether readout and adaptation choices for direct selection remain effective when candidates are scored in separate NLI reranking requests. It combines mean-span pooling with selective adaptation, reporting higher MC11 accuracy with fewer trainable parameters on NeoHorse-1-4B.

  • Problem

    It is unclear whether readout and adaptation choices used for direct selection remain effective when candidates are evaluated in separate NLI requests for reranking.

    Source: §1 · pp. 1–2

  • Method

    The approach combines mean-span pooling over NLI option states with selective backbone adaptation restricted to attention and gated DeltaNet projections.

    Source: §1 · p. 2, §9 · p. 11

  • Results

    On NeoHorse-1-4B, selective adaptation improves MC11 from 50.28% to 52.49%, and adding mean-span pooling reaches 53.39% while reducing trainable parameters by 55.2%.

    Source: §9 · p. 11

  • Takeaways & Limitations

    Both design changes improve MC11 accuracy across all four fold–seed settings, with the combined model using less than half the all-module trainable parameters.

    Source: §5.1 · p. 6, §9 · p. 11

Abstract

from arXiv · show

JEV decision models pair a language model backbone with a lightweight scoring head to select among natural-language options. Direct selection compares candidate answers within one request, whereas entailment reranking evaluates each candidate in a separate natural language inference (NLI) request over fixed relation options. We examine whether the readout and adaptation choices used for direct selection remain effective for cross-request reranking. We combine a mean-span readout with selective backbone adaptation. Mean-span pooling replaces the option-ending pointer with an average of the contextualized states inside each NLI option, without adding head parameters. Selective adaptation retains low-rank updates in attention and gated DeltaNet projections while removing feed-forward adapters. On MC11, which averages accuracy over eleven reranking configurations from seven sources, selective adaptation improves NeoHorse-1-4B accuracy over the all-module baseline while cutting trainable parameters by more than half, and mean-span pooling yields further consistent gains at the same parameter count.

1. Introduction

JEV reranking scores each candidate in a separate NLI request, but readout designs suited to direct selection may not transfer to that setting. The paper combines mean-span pooling with selective backbone adaptation and reports improved MC11 accuracy with fewer trainable parameters.

Motivation

  • Entailment reranking evaluates each answer in a separate NLI request, providing a fixed scoring interface when the number and wording of externally supplied answers vary.The question is the premise, each candidate is a hypothesis, and candidates are compared by entailment probability.Source: §1 · p. 1
  • Because NLI option text is fixed while its contextualized states vary with each candidate, readouts effective for within-request comparison may not transfer to reranking.The paper combines mean-span pooling with selective adaptation to gather information from more positions while updating fewer backbone projections.Source: §1 · p. 2

Approach

  • Mean-span pooling averages contextualized states inside each NLI option, while selective adaptation retains attention and gated DeltaNet updates and removes FFN adapters.The shared affine bilinear head makes pooling an exact average of pre-softmax token scores without adding head parameters.Source: §1 · p. 2

Results

  • On MC11, selective adaptation raises accuracy from 50.28% to 52.49%, and mean-span pooling reaches 53.39% at the same reduced parameter count.Both steps improve all four fold–seed settings; the combined changes reduce trainable parameters by 55.2%.Source: §1 · p. 2
  • Across fixed-checkpoint PIQA and SocialIQA comparisons, the pointer leads on average under direct selection, while pooling leads under entailment reranking.The relative performance of option readouts depends on the scoring interface.Source: §1 · p. 2

2. Readout and Adaptation Design

The design defines reranking as separate NLI requests for each candidate and compares option readouts that use either a closing delimiter or the option span. Selective adaptation limits trainable updates to attention and gated DeltaNet projections.

2.1. Typed requests and candidate reranking

  • For entailment reranking, each candidate is a hypothesis in a separate request with entailment, neutral, and contradiction options; candidates are ranked using entailment probability.The shared option text receives distinct contextual representations because the question and candidate hypothesis precede it.Source: §2.1 · pp. 2–3

2.2. Option-ending and mean-span readouts

  • Mean-span pooling averages the contextualized states of each option’s interior tokens, while the pointer reads its closing delimiter; both use the same head size.The affine projection makes the pooled score an exact average of contextual token scores before the option softmax, and pooling provides a direct score-gradient path through every interior state.Source: §2.2 · pp. 3–4

2.3. Selective adaptation

  • Selective adaptation keeps low-rank updates in attention and gated DeltaNet projections, removing FFN adapters while retaining pretrained FFN weights in the forward computation.The compared scopes cross with both readouts while keeping head size and adapter rank fixed.Source: §2.3 · p. 4

3. Experimental Setup

The experiments evaluate NeoHorse-1-4B reranking with a shared training recipe and matched comparisons across folds and seeds. MC11 averages accuracy across eleven configurations, while additional evaluations test other reranking and decision settings.

Models and training

  • Main experiments use NeoHorse-1-4B with a one-epoch reference recipe and rank-64 adapters; a separate pointer–pooling comparison uses Qwen3-4B.The reference recipe uses a OneCycle schedule with a peak learning rate of 1.25×10−5 for the head and adapters.Source: §3 · p. 4

Training data

  • The training mixture combines web decisions and policy tasks with sentiment, topic, intent, and entailment classification, converting sources to a shared request format.Public-source labels and model-generated reference answers retain each question’s native options and target distribution.Source: §3 · p. 4

Training splits and comparisons

  • Eight training pools contain 17,661 records and 24,001 questions; two folds crossed with two seeds yield four matched settings.Configurations are trained afresh with matched initial tensors, data order, and augmentation streams within each setting.Source: §3 · pp. 4–5

Evaluation and reporting

  • MC11 equally weights reranking accuracy across eleven configurations from seven sources, and additional evaluations include PIQA, CommonsenseQA, JevBench, and Kev.Accuracy is reported in percent; paired changes are calculated within each setting and summarized by their mean and descriptive SE.Source: §3 · p. 5

4. Readout Behavior Across Decision Interfaces

Fixed-checkpoint comparisons on the same PIQA and SocialIQA questions test whether readout performance changes between direct selection and entailment reranking. Pooling is stronger on average within reranking, although the comparison varies with NLI option wording.

Readout comparison

  • On PIQA, the relative readout difference shifts from −0.12 to +8.07 pp between direct selection and reranking, with pooling leading under reranking.Table 1 reports four-setting means and descriptive SEs for the same questions and fixed checkpoints; neither source enters MC11.Source: §4 · p. 5, Table 1 · p. 5
  • Pooling leads the pointer on average within reranking on both PIQA and SocialIQA, while direct selection has higher mean absolute accuracy on both sources.SocialIQA shows the same direction of interface interaction at both training budgets.Source: §4 · p. 6

Option wording

  • Across 18 PIQA combinations of three descriptions and six label orders, pooling’s mean advantage is +3.18 ± 3.27 pp, but one description set favors the pointer by 1.80 pp.The comparison averages over orders for the description-specific result.Source: §4 · p. 6

5. Reranking Accuracy and Parameter Efficiency

The section compares readout and adapter-scope choices under a shared MC11 recipe, then examines how parameter allocation, training choices, and task-level effects shape reranking accuracy.

5.1. Combining Selective Adaptation and Span Pooling

  • 53.39% MC11 accuracy for no-FFN mean-span pooling is the highest of the four matched configurations, versus 50.28% for the all-module pointer, with 55.2% fewer trainable parameters.Removing FFN adapters reduces trainable parameters from 131.17M to 58.82M; pooling adds 0.91 ± 0.35 pp at the same parameter count, and both changes improve every fold–seed setting.Source: Table 2 · p. 6, §5.1 · p. 6, Table 3 · p. 6

5.2. Adapter Scope and Parameter Budget

  • 72.35M trainable parameters are in FFN adapters, and removing them improves the pointer by 2.21 ± 0.66 pp and pooling by 2.39 ± 0.35 pp.At a similar parameter budget, the rank-28 all-module pointer reaches 52.40%, close to the rank-64 no-FFN pointer’s 52.49%.Source: §5.2 · p. 7

5.3. Option Aggregation and Scoring Head

  • 51.00% MC11 for mean-span pooling exceeds the last-content control’s 49.11% by 1.89 ± 0.97 pp, supporting aggregation across the option span over either tested pointer position.Adding a two-layer residual correction yields 50.33% MC11, close to the reference pointer, so the final design retains the bilinear head.Source: §5.3 · p. 7

5.4. Training Recipe and Reference Quality

  • 2.83 ± 0.79 pp is the average MC11 gain from lowering the adapter rate, while changing the head rate has little average effect (+0.02 ± 0.43 pp).Both readouts perform better after one epoch in every setting; the higher-rate head-first schedule reaches 53.54% with a no-FFN pointer, while its combination with pooling remains untested.Source: §5.4 · p. 8

5.5. Task-Level Sources of Improvement

  • About 13.3 pp of final-model advantage on HellaSwag combines roughly 4.2 pp from pooling and 9.1 pp from removing FFN adapters; the final model also improves both ARC and MMLU configurations and WinoGrande.Chess pooling gains rise from 1.95 pp under all-module adaptation to 7.35 pp without FFN adapters; pooling and scope gains remain positive under either weighting scheme when any one source is omitted.Source: §5.5 · p. 8, Figure 3 · p. 9

6. Additional Reranking Sources and Backbone Comparison

The section tests readout and adapter choices on PIQA and CommonsenseQA, then checks mean-span pooling under all-module adaptation on Qwen3-4B. Pooling helps on several sources but does not improve Qwen3-4B’s aggregate MC11 accuracy.

6.1. PIQA and CommonsenseQA

  • Table 5 reports paired four-setting comparisons; its adapter-scope results use lower-rate one-epoch pointers and do not evaluate the final no-FFN pooling model.The readout comparison uses the same model group, while the learning-rate comparison changes head and adapter rates together.Source: Table 5 · p. 9
  • Pooling improves the lower-rate pointer models by 1.26 pp on PIQA and 4.01 pp on CommonsenseQA; omitting FFN adapters also improves both sources in every tested setting.The no-FFN comparison reaches 80.25% on PIQA and 63.80% on CommonsenseQA.Source: Table 5 · p. 9, §6.1 · p. 10

6.2. Readout on Qwen3-4B

  • 51.70% for pooling versus 52.17% for the pointer on Qwen3-4B MC11: task-level pooling gains, including HellaSwag and ARC, are outweighed by declines elsewhere.The four-option GSM8K loss is 4.28 pp.Source: §6.2 · p. 10

7. Trade-offs on Direct Selection

The direct-selection comparison finds that pointer readouts outperform pooling across both adapter scopes on the evaluated suites. No-FFN adaptation raises MC11 over the all-module pointer while maintaining JevBench accuracy, but pooling narrows those gains.

7. Trade-offs on Direct Selection

  • 73.79% JevBench and 78.12% Kev with no-FFN pooling are 3.32 and 3.25 pp below the all-module pointer, respectively.No-FFN pointer reaches 77.52% MC11 versus 77.11% for the all-module pointer.Source: §7 · p. 10, Table 6 · p. 10
  • Across both adapter scopes, the pointer has higher direct-selection accuracy than pooling on these suites; Kev declines include decision-test and transfer subsets.The comparison supports choosing the readout for the intended scoring task, alongside the fixed-checkpoint interface comparisons in Section 4.Source: §7 · p. 10

8. Related Work

Related work covers JEV behavior and scoring interfaces, representation readouts, and adaptation controls. This paper combines readout and adapter-scope comparisons in trained JEV models used for cross-request entailment reranking.

8. Related Work

  • Prior JEV studies report uneven performance across task types, while interventions and scoring interfaces can alter model behavior or multiple-choice scores [24], [1] [47], [34] [53].LLM-as-Jev studies joint suffix scoring and KL-anchored fine-tuning [25].Source: §8 · p. 10
  • Pointer and biaffine architectures use shared projections, while sentence and span representation work studies which token states to aggregate [14], [32, 42], [45] [2].Tang and Yang [40] compare pooling and attention with a fixed embedding model, data, and backbone.Source: §8 · pp. 10–11
  • Prior adaptation research examines module placement, capacity, optimization, and alternatives such as gradual unfreezing and head fitting [18], [22], [5], [15] [8], [17], [20], [50], [13].The paper brings readout design and adaptation scope together in a typed decision model for cross-request entailment reranking.Source: §8 · p. 11

9. Conclusion

The conclusion pairs mean-span pooling with selective adaptation for cross-request entailment reranking and reports improved MC11 accuracy with fewer trainable parameters. The appendices describe evaluation, training, reproduction, and evidence boundaries.

9. Conclusion

  • 53.39% MC11 follows selective adaptation’s 52.49%, up from the 50.28% all-module pointer; both steps are positive in all four fold–seed settings.Together, the changes improve reranking accuracy by about 3.1 pp while reducing trainable parameters by 55.2%.Source: §9 · p. 11

A.1. MC11 populations and scoring

  • MC11 averages eleven configuration accuracies across seven sources, with 46,475 question occurrences per model; some questions recur across configurations.Appendix D examines equal source weighting and individual source omissions.Source: §A.1 · p. 16, Table 7 · p. 16

A.2. Training corpus and grouped folds

  • Training uses a predominantly English corpus of 22,035 records and 29,897 questions, split by linked groups into eight training pools and held-out development, calibration, and test sets.Four settings combine folds 2 and 3 with seeds 44 and 45.Source: §A.2 · pp. 16–17

A.3. Serialization, backbone, and optimization

  • The JEV request serializes state, instruction, and options with explicit delimiters, and training minimizes cross-entropy against target option distributions.The NeoHorse-1-4B setup uses rank-64 LoRA across attention and gated DeltaNet projections, with feed-forward adapters included only in all-module adaptation.Source: §A.3 · pp. 17–18

A.4. Configurations, matching, and reuse

  • Configurations are trained independently with matched starting parameters and data streams where applicable; named comparisons average within-setting score differences across four fold–seed settings.Some studies add models while reusing core reference pointers.Source: §A.4 · p. 18

A.5. Other endpoints

  • JevBench equally weights 18 families, while Kev averages question-level accuracy across development and test partitions of three datasets; fixed-checkpoint interface comparisons use PIQA and SocialIQA.The four reported settings share training pools and evaluation questions, so their standard errors describe variation within this design.Source: §A.5 · p. 18, §A.6 · p. 19

A.6. Evidence coverage and reproduction

  • Evidence boundaries include unassessed semantic paraphrases and pretraining exposure, unavailable framework forks, and source-generation and eligibility steps that remain unreplayed.The original GPQA few-shot configuration draws demonstrations from the Diamond evaluation population; a separate evaluation uses disjoint demonstrations.Source: §A.6 · p. 19

B.1. Additional training comparisons

  • Additional experiments test joint learning-rate changes, schedules, and adapter placement using H100- or H20-trained models and their own matched references.Those lower-rate all-module references reach 50.33% and 49.34% MC11, respectively, versus 50.28% for the core reference.Source: §B.1 · p. 19

B.2. Task effects, adapter count, and backbone

The appendix tests how task, adapter scope, learning rate, schedule, inference setting, and NLI wording affect readout comparisons. Results include task-specific trade-offs, sensitivity across settings, and a residual correction that changes very few trained-model selections.

B.2. Task effects, adapter count, and backbone

  • Removing FFN adapters from the pooling model adds 9.10 pp on HellaSwag and 15.20 pp on chess, but lowers both GSM8K and GPQA configurations.Under the corresponding pointer comparison, scope changes are also largest on HellaSwag and chess, at 9.13 and 9.80 pp.Source: §B.2 · p. 20, Table 8 · p. 20
  • Pooling changes Qwen3-4B MC11 accuracy from 52.17% to 51.70%, a paired difference of −0.47 ± 0.59 pp, with one positive setting.HellaSwag retains a gain, while ARC differences are small and GSM8K declines; no additional-source evaluation was conducted.Source: §B.2 · p. 20

B.3. Learning-rate scaling and staged adaptation

  • Lowering both rates together raises pointer MC11 from 45.90% to 50.33%, while shrinking pooling’s advantage over the same-rate pointer from 3.95 to 1.26 pp.The head and adapter rates change together, so this comparison does not isolate the adapter-rate main effect.Source: §B.3 · p. 21
  • Head-first adaptation reaches 52.28% MC11 and exceeds head-only-tail training by 2.54 ± 1.49 pp; combining head-first with no-FFN reaches 53.54%.The combined scope-by-schedule interaction is −2.41 ± 1.61 pp, and adding pooling to this higher-rate recipe remains untested.Source: §B.3 · p. 21
  • Across a further seven-pair rate comparison, mean MC11 changes by −0.34 pp, while JevBench and Kev means rise by 1.69 and 0.74 pp.Saved initial-tensor identity is not established, and the authors state that the rate result depends on the interval and comparison tested.Source: §B.3 · p. 21

B.4. Additional sources and inference execution

  • Across two inference environments, pointer/pooling accuracy is 65.27/73.12% on H100 and 65.15/73.22% on H20, yet about 2.6% of selections change.The comparisons are therefore kept within one environment; hardware and service environment vary together.Source: §B.4 · pp. 22–23
  • Across 18 NLI templates, the mean pooling effect is +3.18 ± 3.27 pp, while fixed-template means range from −5.51 pp to positive values.The reported sign changes make the default-template gain sensitive to how NLI relations are described.Source: §B.4 · pp. 22–23

C. Residual Corrections and Training Controls

C.2. Frozen branches, repeated training, and divergence

Residual-head comparisons vary with the training recipe, and small contrasts are sensitive to repeated pointer training. The authors report that the operations behind nondeterministic divergence and its contribution to full-training differences remain unresolved.

C.2. Frozen branches, repeated training, and divergence

  • At the higher rate, the frozen-branch model exceeds the pointer by 1.39 ± 0.38 pp on MC11, and the trained residual model adds 2.02 ± 0.97 pp over the frozen branch.At lower rates, core contrasts are small and reference-dependent, while a separate jointly reduced-rate comparison retains a +0.57 ± 0.18 pp residual advantage.Source: §C.2 · p. 23, Table 11 · p. 24
  • Repeating the pointer from the same saved initialization, nominal seed, and recorded data streams can reverse the small residual-versus-pointer MC11 contrast.Short matched probes also produce different gradients and parameters when CUDA determinism is disabled; the responsible operations and their contribution remain unresolved.Source: §C.2 · p. 24
  • During lower-rate residual training, corrected and pre-correction BF16 logits select the same option on all recorded training questions.BF16 addition can round away a nonzero correction; this training trace differs from the FP32 intervention at higher-rate checkpoints.Source: §C.2 · p. 24

D. Aggregation and Reference Sensitivity

This section tests whether benchmark aggregation and reference-run choice alter reported effects. Individual pooling and no-FFN comparisons are more stable across reweightings than interactions or near-equal-arm rankings.

D. Aggregation and Reference Sensitivity

  • MC11 weights eleven configurations equally, while the seven-source alternative first averages configurations within four sources and then weights seven sources equally.The two leave-one-source-out schemes omit one source and average either the remaining configurations or source means.Source: D. Aggregation and Reference Sensitivity · p. 24, Table 12 · p. 24
  • Pooling and no-FFN main comparisons retain positive four-setting means under both weighting rules and every source omission.By contrast, their interaction reverses under chess omission, and no-FFN rank 64 versus all-module rank 28 can change sign under configuration-weighted omission.Source: Table 12 · p. 24, D. Aggregation and Reference Sensitivity · p. 25
  • Longer training lowers both models’ accuracy, with the pointer losing more; ARC and HellaSwag account for about 63% of the wider gap.Under seven-source weighting, the mean budget interaction remains +3.66 pp, though one setting has a −0.20 pp interaction.Source: Table 12 · p. 24, D. Aggregation and Reference Sensitivity · p. 25
  • Replacing the reference pointer with its repeat changes the all-module pooling effect to +0.55 ± 1.11 pp, while no-FFN and rank-28 pointer gains remain positive in all four settings.The latter gains are +2.04 ± 0.63 and +1.95 ± 0.77 pp, respectively.Source: D. Aggregation and Reference Sensitivity · p. 25
  • Pooling and scope restriction also lower Kev under the three-test-subset average, including the decision test, while smaller residual-control contrasts change sign under that weighting.Source: D. Aggregation and Reference Sensitivity · p. 25

Related papers

Loading 2610.35654v1…