Dynamic Decoupled Conditional Advantage (DDCA)

• The paper introduces DDCA as a method that decouples correctness and efficiency by computing separate advantages for binary verifiable rewards and conditional length penalties.
• It computes length advantages only within the subset of correct responses, dynamically scaling penalties using the pass rate as a measure of task difficulty.
• Applied in both sequence-level (RLVR) and token-level (DEPO) settings, DDCA improves reasoning efficiency while maintaining or enhancing accuracy on benchmarks.

Searching arXiv for the specified DDCA papers to ground the article in the cited preprints. Dynamic Decoupled Conditional Advantage (DDCA) is a reinforcement-learning method for efficient reasoning in settings with verifiable rewards. In the formulation introduced in "Think Dense, Not Long: Dynamic Decoupled Conditional Advantage for Efficient Reasoning" (Peng et al., 2 Feb 2026), DDCA is designed for Reinforcement Learning with Verifiable Rewards (RLVR), where a policy generates a group of sampled responses to a prompt and each response receives a binary correctness signal from a programmatic verifier. The method addresses a specific pathology of group-relative optimization: RLVR can elicit strong multi-step reasoning, yet it often encourages overly verbose traces, and naive length penalties can reduce output length at the cost of accuracy. DDCA therefore decouples efficiency optimization from correctness by computing length advantages only within the correct-response cluster and by scaling length pressure dynamically with the group pass rate, used as a proxy for difficulty (Peng et al., 2 Feb 2026). A distinct usage of the same term appears in "Towards Flash Thinking via Decoupled Advantage Policy Optimization" (Tan et al., 17 Oct 2025), where the paper’s DEPO framework is described through a DDCA-like core mechanism based on token-level advantage decoupling, difficulty-aware length penalties, and advantage clipping. Across both usages, the central idea is to penalize inefficient reasoning without corrupting the primary correctness signal.

1. Problem setting and motivating failure modes

DDCA arises in group-based RLVR, where a prompt $q$ is paired with a set of sampled responses $O = \{o_1,\dots,o_N\}$, and each response is scored by a verifiable correctness reward $r^{\text{acc}} \in \{0,1\}$. In prior group-relative methods such as GRPO and RLOO, the advantage is estimated from within-group reward statistics rather than from an explicit value function. The GRPO-style standardized advantage is

$$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$

and the RLOO advantage is

$$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$

The DDCA paper identifies two structural failures in naive length regularization (Peng et al., 2 Feb 2026). The first is Dilution of Length Baseline. If a linear length penalty is mixed directly into a scalar reward,

$$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$

then incorrect responses, which typically have zero reward and no length signal, depress the group baseline. For a group with correct subset $C$ of size $n$ and incorrect subset $W$ of size $N-n$, the group mean becomes

$O = \{o_1,\dots,o_N\}$0

with $O = \{o_1,\dots,o_N\}$1 the mean length among correct responses. The resulting unnormalized advantage for a correct sample is

$O = \{o_1,\dots,o_N\}$2

The comparison is therefore against $O = \{o_1,\dots,o_N\}$3 rather than $O = \{o_1,\dots,o_N\}$4, so the baseline is shrunk by the pass rate. This causes correct but relatively long solutions to be over-penalized (Peng et al., 2 Feb 2026).

The second failure is Difficulty-Penalty Mismatch. A static penalty coefficient $O = \{o_1,\dots,o_N\}$5 does not adapt to prompt difficulty. When the group pass rate $O = \{o_1,\dots,o_N\}$6 is low, correct solutions tend to be longer, so a fixed length penalty can suppress necessary reasoning. When $O = \{o_1,\dots,o_N\}$7 is high, the same penalty may be too weak to remove redundant reasoning on easy instances (Peng et al., 2 Feb 2026).

A closely related diagnosis appears in DEPO, though at token granularity rather than purely sequence granularity. There, a single sequence-level reward is broadcast across all tokens, which can increase the likelihood of redundant late tokens once a response is correct. Pure sequence-level length penalties can also invert gradient direction, producing negative advantages for correct-but-long sequences and positive ones for incorrect sequences under some group configurations (Tan et al., 17 Oct 2025). This suggests that both versions of DDCA are motivated by the same general problem: reward shaping for efficiency can easily damage the learning signal for correctness if it is coupled too early or too globally.

2. Formal construction in the RLVR formulation

In (Peng et al., 2 Feb 2026), DDCA decouples correctness and efficiency at the advantage level rather than by combining them into a single scalar reward. Correctness remains a binary verifiable signal,

$O = \{o_1,\dots,o_N\}$8

For each prompt, the method forms the correct cluster

$O = \{o_1,\dots,o_N\}$9

It then computes length statistics only within that cluster:

$r^{\text{acc}} \in \{0,1\}$0

Each correct response receives a normalized length score,

$r^{\text{acc}} \in \{0,1\}$1

Because the transformation is bounded, longer-than-peer correct responses obtain a value closer to $r^{\text{acc}} \in \{0,1\}$2, while shorter-than-peer correct responses obtain a smaller value, with boundedness used for gradient stabilization (Peng et al., 2 Feb 2026).

The conditional RLOO length advantage is then defined only inside the correct cluster:

$r^{\text{acc}} \in \{0,1\}$3

This construction removes the baseline dilution effect because incorrect samples do not enter the length baseline at all (Peng et al., 2 Feb 2026).

DDCA next introduces dynamic scaling through the instantaneous pass rate,

$r^{\text{acc}} \in \{0,1\}$4

yielding

$r^{\text{acc}} \in \{0,1\}$5

Thus, hard problems with low pass rate receive weak length pressure, whereas easy problems with high pass rate receive stronger pressure (Peng et al., 2 Feb 2026).

Correctness is optimized by a separate leave-one-out term,

$r^{\text{acc}} \in \{0,1\}$6

and the total decoupled advantage is

$r^{\text{acc}} \in \{0,1\}$7

This is the defining mechanism of DDCA in (Peng et al., 2 Feb 2026): correctness and efficiency remain separate until the advantage combination step, and length shaping is conditionally applied only to correct responses.

3. Training objective, estimator structure, and operational procedure

The policy-gradient objective in (Peng et al., 2 Feb 2026) is expressed over full response trajectories:

$r^{\text{acc}} \in \{0,1\}$8

Here $r^{\text{acc}} \in \{0,1\}$9 is held fixed when computed with $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$0, and KL regularization constrains deviation from a reference policy (Peng et al., 2 Feb 2026). The corresponding policy gradient is

$$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$1

Variance reduction is obtained from leave-one-out baselines in both the correctness and length components (Peng et al., 2 Feb 2026).

The operational loop is correspondingly simple. For each prompt, the policy samples $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$2 responses, computes binary correctness with a programmatic verifier, forms the correct cluster, computes the pass rate, and obtains $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$3. If $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$4, the algorithm computes in-cluster mean and standard deviation of lengths, converts each correct response length to a sigmoid-bounded score, computes the conditional RLOO length advantage, scales it by $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$5, and sets the length term to zero for incorrect samples. If $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$6, the length term is disabled for stability. The total advantage is then used in a PPO- or RLOO-style update with KL to a reference policy (Peng et al., 2 Feb 2026).

The paper emphasizes several operational details. Correctness is verified by task-specific automatic checkers, such as boxed-answer parsing with exact or numeric equality for math and unit tests for code, though the experiments focus on math. Group size is $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$7, with $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$8 used during training. The pass rate $$\hat{A}_i = \frac{r(o_i) - \mu}{\sigma}, \quad \mu = \frac{1}{N}\sum_{j=1}^N r(o_j), \quad \sigma = \sqrt{\frac{1}{N}\sum_{j=1}^N (r(o_j) - \mu)^2 + \epsilon},$$9 is computed on the fly with no smoothing reported. A small $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$0 is added to $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$1 to avoid division by zero, and when $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$2 the method sets $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$3 for all samples (Peng et al., 2 Feb 2026).

The computational overhead is limited. Computing $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$4, $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$5, and $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$6 is $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$7 per prompt, memory overhead consists mainly of storing lengths and binary correctness flags, and runtime overhead is reported as negligible relative to generation (Peng et al., 2 Feb 2026).

4. Token-level DDCA in DEPO

A different instantiation of the DDCA idea appears in "Towards Flash Thinking via Decoupled Advantage Policy Optimization" (Tan et al., 17 Oct 2025). There, DDCA refers to DEPO’s core mechanism: an advantage-decoupled algorithm, a difficulty-aware length penalty, and an advantage clipping scheme. The setting remains grouped policy optimization, but the decoupling operates at the token level rather than only across whole responses.

The DEPO formulation decomposes a generated response into an efficient segment and an inefficient segment. A generative reward model (GRM), fine-tuned from Qwen2.5-Instruct-7B, is trained to detect the first sentence in the chain of thought that reaches the correct answer, denoted by $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$8. The efficient segment is $$A_i^{\text{RLOO}} = r(o_i) - \frac{1}{N-1}\sum_{j\neq i} r(o_j).$$9 and the inefficient segment is $$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$0 (Tan et al., 17 Oct 2025). Tokens are masked by

$$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$1

Redundancy within the inefficient segment is scored by

$$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$2

where $$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$3 is the maximum number of detected transition markers associated with alternative reasoning paths and $$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$4 is the count of self-reflection tokens. A rule-based matcher computes these statistics over $$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$5 (Tan et al., 17 Oct 2025).

The dynamic weighting function is then

$$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$6

For correct rollouts, inefficient tokens receive the down-weighted advantage $$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$7, while all other tokens receive $$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$8:

$$R(o_i) = \begin{cases} 1 - \gamma |o_i|, & \text{if } \text{is\_equivalent}(\hat{y}_i,y)=1,\ 0, & \text{otherwise}, \end{cases}$$9

The paper also presents this in decomposed form as efficient and inefficient components of the advantage (Tan et al., 17 Oct 2025).

DEPO combines this token-level decoupling with a difficulty-aware length penalty over correct responses only:

$C$0

where $C$1 is the number of correct rollouts in the group and serves as a difficulty estimate: larger $C$2 implies an easier question and therefore a stronger penalty (Tan et al., 17 Oct 2025). In contrast to (Peng et al., 2 Feb 2026), DEPO mixes this length term into total reward before group normalization, but it then applies sign-conditioned advantage clipping to preserve update direction:

$C$3

The policy objective uses a PPO-style clipped surrogate with the decoupled per-token advantages (Tan et al., 17 Oct 2025).

The two papers therefore use the same broad terminology to describe related but non-identical mechanisms. In (Peng et al., 2 Feb 2026), DDCA is a sequence-level advantage decomposition in RLVR with conditional in-cluster length comparison and pass-rate scaling. In (Tan et al., 17 Oct 2025), DDCA denotes DEPO’s token-level decoupling, difficulty-aware reward shaping, and sign-consistent advantage clipping. A plausible implication is that “DDCA” functions more as a design pattern—decoupled correctness and efficiency signals with difficulty adaptation—than as a single fixed estimator across all papers.

5. Empirical behavior and reported results

The principal empirical evidence for DDCA as a named method comes from (Peng et al., 2 Feb 2026). The paper evaluates two backbones, DeepSeek-R1-Distill-Qwen-1.5B and DeepScaleR-1.5B-Preview, on GSM8K, MATH500, AMC23, and AIME25. The reported pattern is consistent across datasets: DDCA reduces generated tokens while maintaining or improving accuracy relative to adaptive baselines.

For the DeepSeek-R1-Distill-1.5B backbone, the following results are reported (Peng et al., 2 Feb 2026).

Benchmark	Base	DDCA
GSM8K	81.2% @ 1683 tok	85.0% @ 1211 tok
MATH500	85.0% @ 4662 tok	86.2% @ 3248 tok
AMC23	71.8% @ 7945 tok	76.3% @ 5556 tok
AIME25	21.9% @ 12158 tok	26.5% @ 9616 tok

The paper states that on GSM8K DDCA yields a 28% token reduction with a 3.8% accuracy gain relative to base, while on AIME25 it yields a 21% token reduction with a 4.6% accuracy gain. It also notes that the linear-penalty baseline reduced tokens but degraded accuracy, whereas DDCA improved both. The reported AES is best overall at 0.42 for this backbone (Peng et al., 2 Feb 2026).

For the DeepScaleR-1.5B-Preview backbone, the reported results are as follows (Peng et al., 2 Feb 2026).

Benchmark	Base	DDCA
GSM8K	86.9% @ 1700 tok	86.9% @ 672 tok
MATH500	90.1% @ 3093 tok	88.8% @ 2082 tok
AMC23	81.7% @ 4993 tok	83.2% @ 3515 tok
AIME25	31.1% @ 8301 tok	29.8% @ 6109 tok

Here DDCA preserves GSM8K accuracy while reducing tokens by 60.5%, and on AIME25 it reduces tokens by 26% with a 1.3% accuracy decrease that the paper describes as “comparable accuracy” with substantial efficiency gains. The reported AES is again best overall, at 0.30 (Peng et al., 2 Feb 2026).

The ablation results in (Peng et al., 2 Feb 2026) are especially important for interpreting the method. On MATH, described as easy in the ablation discussion, removing dynamics (“Static”) yields the shortest outputs but harms hard tasks; removing decoupling (“Coupled”) produces longer outputs than DDCA. On AIME25, described as hard, “Static” collapses from 26.5% to 21.9% accuracy while shortening outputs further, and “Coupled” yields worse efficiency than DDCA. This supports the claim that both conditional decoupling and difficulty-aware scaling are functional parts of the reported performance.

The DEPO paper reports a related but broader efficiency effect under its DDCA terminology. On DeepSeek-R1-Distill-Qwen-7B, average accuracy is 71.1% versus 69.3% for the original model, with average length 4656 versus 7591, described as approximately 38.7% shorter. On the 1.5B model, average accuracy is 56.1% versus 54.0% original, with average length 5510 versus 9048, described as approximately 39.1% shorter (Tan et al., 17 Oct 2025). It also reports that overlong responses are reduced from approximately 10.7% of rollouts under GRPO to approximately 0.1% under DDCA, and that redundant reasoning steps in inefficient segments are reduced by about half. Since DEPO’s estimator differs from (Peng et al., 2 Feb 2026), these results should not be treated as measurements of the same algorithmic object, but they do indicate empirical support for the broader decoupled-advantage approach.

6. Implementation details, guidance, and edge cases

The implementation details reported in (Peng et al., 2 Feb 2026) are specific. The backbones are DeepSeek-R1-Distill-Qwen-1.5B and DeepScaleR-1.5B-Preview. Training data consist of 2470 problems from AIME/AMC with Prime/Process RL preprocessing. Evaluation is conducted on GSM8K, MATH500, AMC23, and AIME25. Inference uses temperature 0.6, top-p 0.95, and context up to 16K tokens. Evaluation rollouts are 4 for GSM8K and MATH500 and 16 for AMC23 and AIME25 (Peng et al., 2 Feb 2026).

Training uses the Verl (HybridFlow) and slime codebases on 4× RTX-5090 on one node. Generation during RL has maximum response length 8192, group size $C$4 responses per prompt, 8 prompts per iteration, and global batch size 32. The optimizer is policy gradient with PPO-style infrastructure and KL to reference, with learning rate $C$5 and KL coefficient $C$6. The prompt template is reported verbatim as:

“$$QUESTION. Please reason step by step, and put your final answer within \boxed{.}”</p> </blockquote> <p>The principal DDCA hyperparameter is$$C$7, the global efficiency strength, tuned in a sensitivity study over values such as 0.2–0.5. Larger $C$8 produces more token savings but can reduce hard-task accuracy. The pass-rate scaling function is fixed as $C$9 with no smoothing reported. Conditional normalization is based on within-cluster z-scores and sigmoid bounding, and the method uses the practical safeguard $n$0 whenever $n$1 (Peng et al., 2 Feb 2026).

The practical guidance given in the paper is correspondingly targeted. Existing RLVR loops such as GRPO or RLOO can retain their standard structure: sample grouped responses, verify correctness, compute the accuracy advantage, build the correct cluster, compute the conditional length branch when $n$2, scale by $n$3, and combine the two advantages before the PPO/RLOO update with KL to reference. For easy datasets such as GSM8K, the paper states that $n$4 can be higher because DDCA will push strong compression while preserving accuracy; for hard datasets such as AIME25, it recommends moderate $n$5 and reliance on $n$6 to relax penalties (Peng et al., 2 Feb 2026).

Several edge cases are explicitly noted. If $n$7 or $n$8, all $n$9 values should be set to zero and only correctness should be optimized. If $W$0, the length signal is strong, and the paper suggests considering a cap on $W$1 or on $W$2 magnitude if over-compression is observed. For noisy verifiers, the recommendation is to use strict parsers for boxed answers, post-hoc numeric normalization, or reliable unit tests. The paper also notes that small $W$3 can yield high variance and suggests, as possible stability aids not used in the paper, increasing $W$4 or smoothing $W$5 with a moving average (Peng et al., 2 Feb 2026). Because these latter suggestions are framed as possible stability aids and explicitly marked as not used in the paper, they should be read as practical extrapolations rather than as part of the canonical method.

7. Relation to prior methods, limitations, and interpretive issues

DDCA is positioned against several classes of prior methods. In (Peng et al., 2 Feb 2026), coupled linear penalties, ThinkPrune, and non-linear or adaptive penalties such as TLMRE are grouped together as methods that still mix length with correctness in a single reward and compute baselines across all samples. According to the paper, they therefore retain the structural issues of baseline dilution and static or misaligned penalty strength across hard and easy prompts. DDCA’s contribution is instead to operate at the advantage level, first by conditioning length comparison on the correct cluster and second by dynamically scaling length pressure with pass rate (Peng et al., 2 Feb 2026).

The DEPO paper makes a similar comparison but emphasizes token-level behavior. PPO-LM and GRPO broadcast the same sequence-level advantage to all tokens, thereby reinforcing redundant late reasoning once a response is correct. Length-penalized RL without decoupling can degrade accuracy because correct-but-long sequences receive negative advantages globally. In contrast, DEPO’s DDCA selectively down-weights inefficient tokens only when the rollout is correct and preserves positive updates on efficient tokens via sign-consistent advantage clipping (Tan et al., 17 Oct 2025).

Several limitations are stated directly in (Peng et al., 2 Feb 2026). The pass rate $W$6 is only a local and noisy estimate of difficulty, and can fluctuate when $W$7 is small. Although effective empirically, the paper notes that future work could explore smoothed or learned difficulty signals. The method also does not eliminate the possibility that useful long reasoning may be pruned if $W$8 is too large, especially when verbosity is genuinely necessary for verification. The paper further notes concerns about chain-of-thought suppression: DDCA is intended to preserve depth on hard problems by scaling length pressure down when pass rate is low, but aggressive settings could still suppress beneficial exploration, so monitoring pass@K is recommended. Finally, the method is demonstrated on binary verifiable tasks, especially math; extension to graded or non-binary verifiers, or to open-ended generation, would require redesign of correctness and in-cluster normalization (Peng et al., 2 Feb 2026).

A terminological caveat is also necessary. The phrase “Dynamic Decoupled Conditional Advantage” names the method in (Peng et al., 2 Feb 2026), but (Tan et al., 17 Oct 2025) uses DDCA to refer to DEPO’s internal mechanism of advantage decoupling, difficulty-aware length penalty, and advantage clipping. The two share conceptual commitments but differ in estimator structure, granularity, and training pipeline. This suggests that the literature should not treat all mentions of “DDCA” as interchangeable without reference to the specific paper. In (Peng et al., 2 Feb 2026), DDCA denotes a conditional sequence-level length-advantage branch within RLVR. In (Tan et al., 17 Oct 2025), it denotes a token-level decoupled advantage mechanism embedded inside DEPO. Both, however, exemplify a common research direction: reasoning efficiency should be optimized conditionally and adaptively rather than by imposing a uniform length penalty on all outputs.