TACLer: Tailored Curriculum Reinforcement Learning

Updated 5 February 2026

The paper introduces an innovative curriculum reinforcement learning strategy that dynamically schedules training difficulty, improving complex reasoning accuracy by over 9%.
It implements a dual-mode approach blending explicit chain-of-thought with direct answering, achieving up to 50% computational savings.
Empirical results on math benchmarks demonstrate significant gains including +11.2% accuracy and nearly 50% token reduction compared to baseline models.

TACLer refers to a distinct suite of methodologies and systems unified by the concept of curriculum learning, broadly encompassing advanced strategies for data complexity management in both LLM training and psychologically-informed multi-agent simulation. The TACLer designation appears in contemporary research as a label for model-tailored curriculum practices in reinforcement learning for complex reasoning tasks (Lai et al., 29 Jan 2026), as a multi-agent architecture for transactional analysis simulation—colloquially referred to in practice as the "TACLer" system (Zamojska et al., 19 Oct 2025)—and, less formally, as a threshold-adaptive curriculum learning strategy in medical NLP (Ren et al., 17 Oct 2025). This article focuses primarily on the TACLer methodology as introduced in "TACLer: Tailored Curriculum Reinforcement Learning for Efficient Reasoning" (Lai et al., 29 Jan 2026), but provides relevant context for its connections to the broader TACL and TACLA frameworks.

1. Definition and Motivation

TACLer is a model-tailored curriculum reinforcement learning (RL) framework for LLMs in complex reasoning domains. Its central innovation lies in dynamic scheduling of training data difficulty based on instantaneous model proficiency, and in hybridizing explicit chain-of-thought (CoT) reasoning with concise, direct-answer generation ("NoThinking" mode). The resultant training recipe yields substantial reductions in computation—over 50% relative to long-CoT RL—and token usage, while delivering improvements in solution accuracy for mathematical reasoning benchmarks exceeding 9% over the base models (Lai et al., 29 Jan 2026). These principles are distinct from static curricula and directly address computational inefficiency and overthinking endemic to conventional RL-based CoT prompting.

2. Tailored Curriculum Learning

In TACLer, curriculum difficulty is not adjudicated by input length or handcrafted heuristics. Instead, at each RL stage, the current policy's ability to solve each training example is evaluated:

Easy tier (G1): Model produces a complete CoT or NoThinking response and the answer is correct.
Medium tier (G2): Model completes the response but the answer is incorrect.
Hard tier (G3): Model fails to generate a complete solution within the maximum allowed context (truncation).

The curriculum proceeds across three RL stages. At the end of each stage, the entire training dataset is relabeled according to the updated policy's performance, and the next stage increases the mixture proportion of harder instances. Early stages review and reinforce what the learner already knows, while later stages progressively emphasize unresolved or failure cases. The formal scheduling process is summarized in the algorithm below:

Input: Base policy π₀, training set D
for stage t in {1,2,3} do
    # 1. Difficulty labeling
    Partition D into G1, G2, G3 by running π_{t-1} on each q∈D

    # 2. Construct curriculum mixture
    if t==1:
        S ← sample_from(G1∪G2)           # no G3 in stage1
    else if t==2:
        S ← mixture({G1:50%, G2:15%, G3:35%})
    else:
        S ← D                            # full data in stage3

    # 3. RL training on S to obtain π_t
    π_t ← GroupRelativePolicyOptimisation(π_{t-1}, S)
end for
Output: π₃ as final TACLer policy

This iterative re-labeling is robust, leveraging the model's self-assessed weaknesses to guide the data presentation order.

3. Hybrid Thinking / NoThinking Reasoning Paradigm

TACLer implements a dual-headed reasoning strategy:

Thinking Mode: Prompts explicitly request a chain-of-thought rationale, enclosed within <thinking>…</thinking> tags, followed by a final answer. This mode favors interpretability and robustness.
NoThinking Mode: Prompts cue the model to solve the problem directly—<thinking> Okay, I think I can solve it directly. </thinking>—suppressing intermediary reasoning steps and producing only the concise answer.

Both modes are trained concurrently (interleaved 50/50 per RL batch), sharing model weights but differing only in textual prefix and targets. At inference, users may select either mode globally; the model, having been exposed to both during training, performs efficiently in both regimes. There is no dynamic per-example switching during training or inference—such an extension is suggested as a future direction.

4. Reinforcement Learning Formulation

TACLer leverages Group Relative Policy Optimisation (GRPO) for its policy updates:

Policy: $\pi_\theta(\mathbf{o}\mid q)$ , where the output $\mathbf{o}$ can be a full CoT or direct answer.
Rewards: Binary, based on answer correctness only.
Rollouts: For each instance, $G$ response rollouts are sampled under the fixed old policy, with standardised advantage assigned per rollout.
Objective:

$J(\theta) = \mathbb{E}_{q,\{o_i\}} \frac{1}{G} \sum_{i=1}^G \min\left( r_i(\theta),\,c_i(\theta) \right)$

where $r_i$ , $c_i$ incorporate the advantage and a ratio clipping heuristic without KL penalty (asymmetric clipping with $\epsilon_{\rm low}=0.2$ , $\epsilon_{\rm high}=0.28$ ). This approach precipitates effective learning and compressed solution traces without instabilities from over-regularization.

Training employs a fixed architecture (DeepSeek-R1-Distill-Qwen-1.5B), 8K token contexts, batch size 128, and standard optimizer parameters. Stagewise training comprises 280 steps (stage 1), 350 steps (stage 2), and 1250 steps (stage 3), notably more efficient than DeepScaleR or FastCuRL comparators (Lai et al., 29 Jan 2026).

5. Empirical Performance and Analysis

Experiments were conducted on the DeepScaleR dataset (40K competition math problems), with evaluation on MATH500, AMC, and AIME 2024/2025. Performance was rigorously compared to both Thinking-mode and NoThinking-mode baselines.

Thinking mode: TACLer achieves +11.2% accuracy over the base R1-Qwen(Thinking), with a 42.7% reduction in response length, and outpaces DeepScaleR by up to 1.7% on certain benchmarks.
NoThinking mode: TACLer achieves +9.5% accuracy over R1-Qwen(NoThinking), with a nearly 50% reduction in token length and large gains over other concise-answer baselines (e.g., AutoThink, OverThink).
Hybrid vs. Pure: Hybrid training produces nearly equal accuracy as pure NoThinking but with shorter responses, indicating compression of capabilities.
Training efficiency: TACLer achieves >50% computational savings by reducing average rollout lengths and truncation rates (Figure 1, (Lai et al., 29 Jan 2026)).

A summary table organizes salient results:

Mode	Accuracy Gain vs. Base	Avg. Token Reduction	Notable Benchmark Lead
Thinking	+11.2%	–42.7%	+1.7% (AIME2024 vs. DeepScaleR)
NoThinking	+9.5%	–49.3%	+6.4% (AIME2024 vs. AutoThink)

6. Extensions, Limitations, and Theoretical Significance

The curriculum's principal innovation is its reliance on the model's own capability judgments to determine difficulty progression, rather than externally imposed heuristics or static measures. Hybrid RL on both CoT and concise-answer modes produces policies simultaneously interpretable and efficient.

Identified limitations include the requirement for repeated full-dataset inference-based relabeling between RL stages (~5 GPU-hours per cycle), and a lack of dynamic instance-level mode selection at inference. Proposed extensions are automatic gating for on-the-fly mode selection, expansion to multi-step domains such as code synthesis or planning, and tighter curriculum-reward coupling.

A plausible implication is that model-tailored curricula may generalize to broader RL domains, potentially yielding similar gains in computational and reasoning efficiency when adapted to other forms of multi-stage reasoning or structured prediction.

7. TACLer in the Broader Methodological Landscape

The TACLer approach stands alongside and conceptually intersects with other advanced curriculum mechanisms, notably:

Threshold-Adaptive Curriculum Learning (TACL) in Medical NLP: Employs dynamic progression based on thresholded macro-F1 improvements, clustering schema on contextualized embeddings, and pre-defined data partitions in multilingual clinical applications (Ren et al., 17 Oct 2025). TACL’s threshold adaptation uses the criterion $\gamma_{\Delta} < \beta\bar\gamma$ to govern stage transitions, with empirical successes in ICD coding, readmission prediction, and syndrome differentiation.
TACLA Multi-Agent Social Simulation: Under the moniker "TACLer," indicates a distinct area, leveraging orchestrated LLM agents to simulate transactional analysis by dynamic ego-state activation within psychologically-grounded response protocols (Zamojska et al., 19 Oct 2025).

These frameworks collectively reflect a contemporary interest in adaptive curricula—in RL or supervised setups, for reasoning or for simulating humanlike learning and behavior—as a cornerstone for scalable, efficient, and modular model training in language and sequential decision problems.