GateKD: Confidence-Gated Closed-Loop Distillation for Robust Reasoning

Kasidit Sermsri; Teerapong Panboonyuen

arxiv: 2605.13136 · v2 · pith:BSYSUPX3new · submitted 2026-05-13 · 💻 cs.CL

GateKD: Confidence-Gated Closed-Loop Distillation for Robust Reasoning

Kasidit Sermsri , Teerapong Panboonyuen This is my paper

Pith reviewed 2026-06-30 21:45 UTC · model grok-4.3

classification 💻 cs.CL

keywords reasoning distillationclosed-loop distillationconfidence gatingknowledge distillationlanguage modelslogical reasoninghallucination reduction

0 comments

The pith

GateKD uses teacher confidence to create a closed feedback loop that selectively distills reliable reasoning into smaller models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Open-loop distillation methods transfer reasoning from large language models but risk passing along noisy or incorrect intermediate steps because they treat all teacher outputs as equally reliable. GateKD counters this by making the teacher a dynamic gatekeeper whose confidence scores decide when to apply supervision, align representations, and filter attention patterns. The three mechanisms operate together in a closed loop that continuously adjusts the distillation based on reliability signals. Tests across commonsense, logical, and symbolic benchmarks with T5 and Flan-T5 models of different sizes show consistent gains over standard baselines, with especially clear benefits for logical and symbolic tasks and stability when data is limited. Removing any one gating part produces measurable drops in student performance.

Core claim

GateKD treats the teacher as a dynamic gatekeeper rather than a static oracle and introduces confidence-gated soft supervision, gated hidden-state evolution, and reliability-filtered attention distillation that together form a closed feedback loop in which teacher confidence continuously modulates the distillation process, reducing hallucination transfer and stabilizing student reasoning.

What carries the argument

The confidence-gated closed-loop distillation framework that uses teacher confidence scores to selectively apply soft supervision, align intermediate hidden states, and filter attention patterns.

If this is right

Smaller models obtain higher accuracy on multi-step reasoning tasks by avoiding transfer of unreliable teacher steps.
The approach maintains gains even when the amount of available distillation data is reduced.
Logical and symbolic reasoning tasks receive larger benefits than commonsense reasoning under this selective transfer.
Each of the three gating mechanisms contributes separately, as shown by performance degradation in targeted ablations.
The method works across different backbone sizes based on T5 and Flan-T5 families.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same gating principle could be tested in distillation settings outside language, such as transferring planning abilities in reinforcement learning agents.
If confidence estimation itself can be improved, the closed-loop benefit would likely increase.
Iterative application where the student later acts as a gated teacher might create self-improving reasoning chains.
The framework suggests a general pattern for handling noisy supervision in any teacher-student setup.

Load-bearing premise

That teacher confidence scores serve as a reliable indicator of correct intermediate reasoning steps rather than overconfident errors.

What would settle it

Training student models on the same benchmarks with all gating components disabled and observing no drop in logical or symbolic reasoning accuracy compared with the full GateKD version.

Figures

Figures reproduced from arXiv: 2605.13136 by Kasidit Sermsri, Teerapong Panboonyuen.

**Figure 1.** Figure 1: Overview of the proposed GateKD framework. Given an input, the teacher and student models process the instance in parallel. The teacher produces predictive distributions, hidden states, and attention maps, along with a confidence score estimated via predictive entropy. This confidence signal acts as a unified gating mechanism that selectively controls three distillation pathways: (i) confidence-gated soft … view at source ↗

**Figure 2.** Figure 2: Qualitative comparison of reasoning behaviors on StrategyQA. The teacher model produces a fluent but speculative reasoning trace and incorrectly predicts “Yes,” conflating theoretical possibility with practical feasibility. In contrast, the mentor model emphasizes physical constraints, sustainability, and real-world viability, leading to the correct prediction “No.” GateKD selectively suppresses low-confid… view at source ↗

**Figure 3.** Figure 3: Qualitative comparison on shuffled object tracking. The teacher produces a fluent but incorrect reasoning trace, leading to a wrong final answer. In contrast, the mentor generates a consistent intermediate state transition and arrives at the correct solution. GateKD selectively suppresses unreliable teacher reasoning and reinforces stable mentor supervision via confidence-aware gating. teacher confidence (… view at source ↗

**Figure 4.** Figure 4: Error correction on date understanding. The teacher incorrectly infers the current date and propagates the error through subsequent steps. The mentor exhibits lower predictive entropy and maintains a coherent temporal reasoning chain, resulting in the correct answer. GateKD prioritizes such low-entropy supervision, preventing error amplification during distillation. (batch-relative vs fixed threshold). (4)… view at source ↗

**Figure 5.** Figure 5: Robust arithmetic reasoning on SVAMP. The teacher prematurely collapses the reasoning process, yielding an incorrect solution. The mentor follows a structured algebraic derivation with consistent intermediate steps. GateKD gates intermediate supervision based on confidence, enabling the student to acquire correct symbolic reasoning patterns. Variant CSQA SQA Shuffled Last Letter Vanilla-KD 63.8 57.2 82.6 5… view at source ↗

**Figure 6.** Figure 6: Illustrative failure case motivating confidence-gated distillation. The example shows a simple string reasoning task—concatenating the last letters of each word in “Max Mikey Cynthia Holly.” Although the teacher explicitly enumerates intermediate steps, it produces an incorrect conclusion (“xyab”) due to unstable reasoning. In contrast, the mentor yields a consistent and correct trace (“xyay”). This discre… view at source ↗

read the original abstract

Distilling multi-step reasoning abilities from large language models (LLMs) into compact student models remains challenging due to noisy rationales, hallucinated supervision, and static teacher-student interactions. Existing reasoning distillation methods, including mentor-based approaches, predominantly operate in an open-loop manner, implicitly assuming uniform teacher reliability and consequently propagating erroneous intermediate reasoning. We propose GateKD, a confidence-gated closed-loop distillation framework that enables robust reasoning transfer by treating the teacher as a dynamic gatekeeper rather than a static oracle. GateKD introduces three complementary mechanisms: (i) confidence-gated soft supervision that selectively distills reliable predictive signals, (ii) gated hidden-state evolution that aligns intermediate representations only when teacher confidence is high, and (iii) reliability-filtered attention distillation that preserves stable reasoning structures while suppressing noisy patterns. These components jointly form a closed feedback loop in which teacher confidence continuously modulates the distillation process, reducing hallucination transfer and stabilizing student reasoning. Extensive experiments across commonsense, logical, and symbolic reasoning benchmarks, using T5 and Flan-T5 backbones of varying sizes, demonstrate that GateKD consistently outperforms strong open-loop distillation baselines. Notably, GateKD yields substantial gains in logical and symbolic reasoning, remains robust under low-resource distillation settings, and shows clear performance degradation when any gating component is removed. Our results highlight that confidence-gated closed-loop supervision is critical for building reliable and scalable small reasoning models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

GateKD proposes closed-loop confidence gating for reasoning distillation but the gains rest on an untested assumption about what teacher confidence actually signals.

read the letter

GateKD adds three gating mechanisms—on soft supervision, hidden-state alignment, and attention—to a closed-loop distillation process so the teacher only transfers signals when its is high. The closed-loop framing and the specific combination of gates on multiple representation levels look like the main departure from prior open-loop distillation work.

The paper does a reasonable job laying out the practical problem of hallucinated rationales getting baked into the student and shows that removing any gate hurts performance in the reported ablations. That at least suggests the components interact with the training signal rather than being inert.

The soft spot is the load-bearing assumption that high teacher confidence reliably marks correct intermediate reasoning steps instead of overconfident errors. If the teacher is confident on wrong chains, the gating will still pass bad supervision while the ablations only demonstrate that the gate matters, not that it selects for accuracy. The abstract claims consistent outperformance on logical and symbolic tasks plus robustness in low-resource settings, but without the actual tables, baseline details, or statistical controls it is impossible to tell how large or reliable the gains are. The stress-test concern lands directly here.

This is for researchers working on compact reasoning models and distillation methods. A reader who needs a concrete way to modulate teacher signals would get the method description and the closed-loop idea out of it.

I would send it for peer review so the experiments can be examined in full.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes GateKD, a confidence-gated closed-loop distillation framework for transferring multi-step reasoning abilities from LLMs to compact student models. It introduces three mechanisms—confidence-gated soft supervision, gated hidden-state evolution, and reliability-filtered attention distillation—that modulate distillation based on teacher confidence to reduce hallucination transfer. Experiments across commonsense, logical, and symbolic reasoning benchmarks with T5/Flan-T5 backbones report consistent outperformance over open-loop baselines, gains in logical/symbolic tasks, robustness in low-resource settings, and performance degradation upon ablation of gating components.

Significance. If the empirical results hold under scrutiny, GateKD offers a practical approach to more reliable reasoning distillation by addressing noisy supervision in open-loop methods. The closed-loop design and ablation evidence for gating importance are strengths in an empirical contribution; however, the work lacks machine-checked proofs, parameter-free derivations, or falsifiable predictions beyond the reported benchmarks.

major comments (2)

[Abstract and gating mechanisms description] The central robustness claim—that gating reduces hallucination transfer—rests on the assumption that teacher confidence reliably marks correct intermediate reasoning steps rather than overconfident errors. The ablations demonstrate degradation when gating is removed but provide no direct evidence (e.g., rationale-level error analysis or correlation between confidence and step correctness) that the gate preferentially filters errors. This is load-bearing for interpreting gains as improved reliability rather than incidental effects.
[Experiments section] Without the full experimental tables, baseline details, statistical tests, or quantitative results (noted as unavailable for assessment), it is not possible to verify the claimed 'consistent outperformance,' 'substantial gains,' or low-resource robustness. The soundness of the empirical claims cannot be evaluated from the abstract alone.

minor comments (1)

[Method] Clarify the precise definition of 'closed-loop' feedback, as the mechanisms appear to use one-way confidence modulation rather than iterative teacher-student updates.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major point below and indicate where revisions will be made.

read point-by-point responses

Referee: [Abstract and gating mechanisms description] The central robustness claim—that gating reduces hallucination transfer—rests on the assumption that teacher confidence reliably marks correct intermediate reasoning steps rather than overconfident errors. The ablations demonstrate degradation when gating is removed but provide no direct evidence (e.g., rationale-level error analysis or correlation between confidence and step correctness) that the gate preferentially filters errors. This is load-bearing for interpreting gains as improved reliability rather than incidental effects.

Authors: We acknowledge that the manuscript does not include direct rationale-level error analysis or explicit correlation between teacher confidence and step correctness. The ablations demonstrate that removing the gating components degrades performance, supporting their contribution, but this is indirect evidence. We will add a new analysis section correlating teacher confidence with intermediate step accuracy to strengthen the robustness interpretation. revision: yes
Referee: [Experiments section] Without the full experimental tables, baseline details, statistical tests, or quantitative results (noted as unavailable for assessment), it is not possible to verify the claimed 'consistent outperformance,' 'substantial gains,' or low-resource robustness. The soundness of the empirical claims cannot be evaluated from the abstract alone.

Authors: The full manuscript contains the complete experimental tables, baseline details, and quantitative results with statistical comparisons in the Experiments section. We will verify that all tables, low-resource results, and significance tests are fully accessible and clearly formatted in the revised version. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method proposal without derivation chain

full rationale

The paper presents GateKD as an empirical distillation framework with three gating mechanisms defined directly from teacher confidence scores. No equations, derivations, or parameter-fitting steps are described that would reduce claimed performance gains to quantities defined by the inputs themselves. The work relies on benchmark experiments rather than any self-referential mathematical construction, self-citation load-bearing premise, or fitted-input prediction. The central assumption about confidence indicating correctness is an empirical hypothesis open to falsification, not a definitional tautology.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Based solely on the abstract, no explicit free parameters, background axioms, or new invented entities are described. The framework builds on standard knowledge distillation concepts but does not detail implementation thresholds or assumptions.

pith-pipeline@v0.9.1-grok · 5784 in / 1075 out tokens · 34159 ms · 2026-06-30T21:45:37.948301+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

3 extracted references · 3 canonical work pages · 1 internal anchor

[1]

arXiv preprint, arXiv:2405.19842

Improve student’s reasoning generalizabil- ity through cascading decomposed cots distillation. arXiv preprint, arXiv:2405.19842. Yuntian Deng, Kiran Prasad, Roland Fernandez, et al

work page arXiv
[2]

arXiv preprint arXiv:2410.09037 , year=

Implicit chain of thought reasoning via knowl- edge distillation. InArXiv preprint. Nathan Ho, Madhu Shridhar, et al. 2023. Teaching small language models to reason. InArXiv preprint. Alex Hsieh et al. 2023. Rationale-based knowledge dis- tillation for neural reasoning.Artificial Intelligence Review. Review of rationale distillation techniques for LLMs. H...

work page arXiv 2023
[3]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Self-consistency improves chain of thought reasoning in language models.arXiv preprint, arXiv:2203.11171. Published at ICLR 2023. Jason Wei, Xuezhi Wang, Dale Schuurmans, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in Neural Information Processing Systems. Shicheng Xu, Liang Pang, Yunchang Zhu, et al. 2025. ...

work page internal anchor Pith review Pith/arXiv arXiv 2023

[1] [1]

arXiv preprint, arXiv:2405.19842

Improve student’s reasoning generalizabil- ity through cascading decomposed cots distillation. arXiv preprint, arXiv:2405.19842. Yuntian Deng, Kiran Prasad, Roland Fernandez, et al

work page arXiv

[2] [2]

arXiv preprint arXiv:2410.09037 , year=

Implicit chain of thought reasoning via knowl- edge distillation. InArXiv preprint. Nathan Ho, Madhu Shridhar, et al. 2023. Teaching small language models to reason. InArXiv preprint. Alex Hsieh et al. 2023. Rationale-based knowledge dis- tillation for neural reasoning.Artificial Intelligence Review. Review of rationale distillation techniques for LLMs. H...

work page arXiv 2023

[3] [3]

Self-Consistency Improves Chain of Thought Reasoning in Language Models

Self-consistency improves chain of thought reasoning in language models.arXiv preprint, arXiv:2203.11171. Published at ICLR 2023. Jason Wei, Xuezhi Wang, Dale Schuurmans, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models.Advances in Neural Information Processing Systems. Shicheng Xu, Liang Pang, Yunchang Zhu, et al. 2025. ...

work page internal anchor Pith review Pith/arXiv arXiv 2023