DiSCTT: Consensus-Guided Self-Curriculum for Efficient Test-Time Adaptation in Reasoning

Mohammad Mahdi Moradi; Sudhir Mudur

cs.CLMar 5, 2026

DiSCTT: Consensus-Guided Self-Curriculum for Efficient Test-Time Adaptation in Reasoning

Authors: Mohammad Mahdi Moradi, Sudhir Mudur

Organizations: Department of Compute Science, Concordia University, Montreal, Canada.

Abstract

Test-time adaptation offers a promising avenue for improving reasoning performance in large language models without additional supervision, but existing approaches often apply a uniform optimization objective across all inputs, leading to inefficient or unstable adaptation on heterogeneous reasoning problems. We propose DiSCTT, a difficulty-aware, consensus-guided self-curriculum framework that dynamically allocates test-time optimization strategies based on instance-level epistemic uncertainty estimated from agreement among sampled reasoning trajectories. Inputs with high consensus are consolidated via supervised fine-tuning using majority-agreed solutions as pseudo-labels, while low-consensus inputs are optimized via reinforcement learning with a consensus-regularized objective that encourages diversity under relevance constraints. Across a broad suite of mathematical and general reasoning benchmarks, DiSCTT consistently outperforms strong test-time adaptation baselines, achieving higher accuracy with reduced variance and substantially lower computation and wall-clock training times. These results demonstrate that explicitly accounting for instance difficulty and uncertainty enables more stable, efficient, and effective test-time adaptation for reasoning models.

Figures & tables

Figure 1: Accuracy on MATH-500 over test-time training updates for different adaptation strategies. Uniform test-time training baselines show early saturation or unstable gains, whereas DiSCTT achieves more stable and sustained improvements through difficulty-aware routing.

Figure 2: AMC accuracy for different adaptation strategies. Difficulty-aware routing and correctness-gated exploration improve convergence stability and prevent performance collapse observed under uniform RL.

Figure 3: Overview of DiSCTT. For each input, multiple reasoning completions are sampled and evaluated for consensus. High-consensus inputs are routed to supervised fine-tuning using majority-agreed solutions as pseudo-labels, while low-consensus inputs are optimized via reinforcement learning. This difficulty-aware bifurcation is periodically recomputed, yielding a self-evolving test-time curriculum.

Model	Method	AMC	MATH-500	AIME-2024	GPQA	HotpotQA	MMLU	Avg.
Qwen-2.5-0.5B-Instruct	Base	$12.2_{\pm 1.21}$	$27.6_{\pm 0.22}$	$1.4_{\pm 1.89}$	$14.1_{\pm 0.32}$	$1.2_{\pm 0.15}$	$36.5_{\pm 0.13}$	15.5
	TTRL	$20.3_{\pm 1.28}$	$37.2_{\pm 0.25}$	$1.5_{\pm 1.66}$	$16.2_{\pm 0.35}$	$10.3_{\pm 0.17}$	$40.2_{\pm 0.16}$	10.9
	EVOL-RL	$23.1_{\pm 1.12}$	$40.0_{\pm 0.21}$	$2.5_{\pm 2.10}$	$18.2_{\pm 0.31}$	$13.7_{\pm 0.13}$	$43.7_{\pm 0.14}$	23.5
	DiSCTT (Ours)	$27.7_{\pm 1.11}$	$49.8_{\pm 0.23}$	$2.9_{\pm 2.33}$	$24.2_{\pm 0.33}$	$18.7_{\pm 0.14}$	$49.4_{\pm 0.12}$	28.7
LLaMA-3.2-1B-Instruct	Base	$15.4_{\pm 1.87}$	$14.0_{\pm 0.24}$	$0.1_{\pm 1.48}$	$7.1_{\pm 0.34}$	$11.7_{\pm 0.16}$	$14.9_{\pm 0.10}$	10.5
	TTRL	$22.5_{\pm 1.54}$	$28.0_{\pm 0.21}$	$1.1_{\pm 1.52}$	$10.6_{\pm 0.31}$	$23.3_{\pm 0.18}$	$29.2_{\pm 0.14}$	19.1

Table 1: Mean accuracy (

\pm

standard deviation) across six mathematical and general reasoning benchmarks under test-time adaptation. Results are averaged over independent runs with different random seeds. DiSCTT consistently outperforms the Base model and prior test-time adaptation baselines across model families and scales. Bold denotes the best-performance for each benchmark and model size.

Figure 4: Out-of-distribution (OOD) reasoning performance under test-time adaptation. DiSCTT consistently outperforms baselines, indicating improved robustness to distribution shift without sacrificing in-distribution performance.

Figure 5: Difficulty-level accuracy dynamics on MATH-500 under different training paradigms. We compare (a) SFT-only, (b) RL-only (GRPO), and (c) DiSCTT across five difficulty levels (L1–L5) over adaptation steps.

Dataset	SFT (%)	RL (%)
AMC	25.0	75.0
MATH-500	47.0	53.0
AIME-2024	3.3	96.7
GPQA	28.8	71.2
HotpotQA	58.3	41.7
MMLU	67.1	32.9

Table 2: Average percentage of instances routed to supervised fine-tuning (SFT) and reinforcement learning (RL) across models for each dataset under consensus-based difficulty routing.

Model	Approach	Dataset	SFT	RL	Total	Cost Ratio
Llama-3.2-1B-Instruct	DiSCTT	MMLU	$2.03\times 10^{16}$	$47.06\times 10^{18}$	$47.08\times 10^{18}$	0.544 ( +45.6% )
Llama-3.2-1B-Instruct	TTRL	MMLU	0	$86.44\times 10^{18}$	$86.44\times 10^{18}$	0.544 ( +45.6% )
Qwen3-4B-Base	DiSCTT	MATH-500	$5.77\times 10^{16}$	$22.41\times 10^{18}$	$22.46\times 10^{18}$	0.683 ( +31.7% )
Qwen3-4B-Base	TTRL	MATH-500	0	$32.78\times 10^{18}$	$32.78\times 10^{18}$	0.683 ( +31.7% )
Qwen2.5-7B-Instruct	DiSCTT	AMC	$4.45\times 10^{15}$	$8.50\times 10^{18}$	$8.50\times 10^{18}$	0.573 ( +43.7% )
Qwen2.5-7B-Instruct	TTRL	AMC	0	$14.05\times 10^{18}$	$14.05\times 10^{18}$	0.573 ( +43.7% )

Table 3: Compute efficiency comparison between DiSCTT and TTRL. Tables present (a) total training compute in FLOPs and (b) wall-clock training time. By routing high-consensus instances to SFT and applying RL only to low-consensus instances, DiSCTT substantially reduces compute relative to RL-only TTRL. Cost Ratio denotes relative compute cost (DiSCTT/TTRL).

Dataset	Accuracy (%)
- Qwen3-4B-Base	51.4
- Correctness gate	62.8
- Population-relative novelty	73.4
- Relevance-aware semantic gating	75.0

Table 4: Impact of reward term decomposition on model performance for MATH-500 dataset

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Model	MATH-500	AIME-2024	ACM	GPQA	HotpotQA	MMLU
Qwen3-1.7B-Base	76.2%	50.0%	80.0%	68.3%	80.1%	70.6%

Appendix

Table 5: Percentage of instances where the majority-selected answer matches the ground truth across reasoning benchmarks

Test-Time Adaptation

Test-time adaptation updates or recalibrates pretrained models at inference to handle distribution shifts, often without target labels.

See all 307 papers

TTCS: Test-Time Curriculum Synthesis for Self-Evolving
Jan 30, 2026Chengyi Yang, Zhishang Xiang, Yunbo Tang +5Test-Time Training Language Model Self-Improvement
Xiamen University · Washington University in St. Louis · Renmin University of China
Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning
Sep 10, 2026Bincheng Gu, Min Gao, Zongwei Wang +3Test-Time Adaptation LLM Reliability
Key Laboratory of Dependable Service Computing in Cyber Physical Society, Chongqing University, China · King’s College London, London, UK · Griffith University, Brisbane, Australia
Query-Conditioned Test-Time Self-Training for Large Language Models
May 13, 2026Chaehee Song, Minseok Seo, Yeeun Seong +2Test-Time Adaptation Test-Time Optimization
School of Electrical Engineering, KAIST · 2Graduate School of Green Growth and Sustainability, KAIST

Jan 30, 2026J/K move · Enter open · S save

TTCS: Test-Time Curriculum Synthesis for Self-Evolving

Chengyi Yang, Zhishang Xiang, Yunbo Tang, Zongpei Teng+4

Xiamen University · Washington University in St. Louis · Renmin University of China

Test-Time Training Language Model Self-Improvement Test-Time RL Numerical Reasoning in Language Models

Open paper AskPDF arXiv

Test-Time Training offers a promising way to improve the reasoning ability of large language models (LLMs) by adapting the model using only the test questions. However, existing methods struggle with difficult reasoning problems for two reasons: raw test questions are often too difficult to yield high-quality pseudo-labels, and the limited size of test sets makes continuous online updates prone to instability. To address these limitations, we propose TTCS, a co-evolving test-time training framework. Specifically, TTCS initializes two policies from the same pretrained model: a question synthesizer and a reasoning solver. These policies evolve through iterative optimization: the synthesizer generates progressively challenging question variants conditioned on the test questions, creating a structured curriculum tailored to the solver's current capability, while the solver updates itself using self-consistency rewards computed from multiple sampled responses on both original test and synthetic questions. Crucially, the solver's feedback guides the synthesizer to generate questions aligned with the model's current capability, and the generated question variants in turn stabilize the solver's test-time training. Experiments show that TTCS consistently strengthens the reasoning ability on challenging mathematical benchmarks and transfers to general-domain tasks across different LLM backbones, highlighting a scalable path towards dynamically constructing test-time curricula for self-evolving. Our code and implementation details are available at https://github.com/XMUDeepLIT/TTCS.

Abstract

Test-Time Adaptation

Explore similar work

TTCS: Test-Time Curriculum Synthesis for Self-Evolving

Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning

Query-Conditioned Test-Time Self-Training for Large Language Models