Reasoning Shrinkage: Techniques & Effects

• Reasoning shrinkage is a set of techniques that reallocates probability mass, gradient signals, and token budgets to improve model reliability.
• It is applied in reinforcement learning to stabilize policy gradients through global baseline regularization and variance reduction methods.
• In post-training, it can narrow reasoning diversity by collapsing paths, while chain-of-thought compression selectively preserves critical steps.

Searching arXiv for the cited papers to ground the article. Searching (Han et al., 5 Feb 2026) EBPO reasoning shrinkage RLVR. Searching (Zeng et al., 5 Nov 2025) shrinkage baselines for RLVR. Searching (Nguyen et al., 16 May 2026) coverage shrinkage in reasoning models. Reasoning shrinkage denotes a cluster of phenomena and techniques that regulate how reasoning models allocate probability mass, gradient signal, or token budget. In recent literature, the term appears in at least three distinct but related senses: estimator shrinkage in reinforcement learning with verifiable rewards (RLVR), where local reward baselines are regularized toward global statistics to stabilize policy optimization; coverage shrinkage in post-trained reasoning models, where pass@1 rises while pass@k falls because the model collapses onto fewer reasoning paths; and chain-of-thought compression, where token mass is deliberately reallocated from redundant steps to critical ones through step-level control (Han et al., 5 Feb 2026, Zeng et al., 5 Nov 2025, Nguyen et al., 16 May 2026, He et al., 6 Jul 2025). A broader statistical lineage connects these uses to empirical Bayes, James–Stein estimation, prediction-powered shrinkage, global-local Bayesian priors, and nonlinear shrinkage more generally (Li et al., 20 Feb 2025, Schmidt et al., 2018, Bhattacharya et al., 2012, Bartz, 2016).

1. Meanings and scope of the term

In RLVR, reasoning shrinkage refers to shrinking noisy prompt-local reward statistics toward a global prior so that policy-gradient updates remain informative under sparse rewards and small rollout groups. This usage is explicit in Empirical Bayes Policy Optimization (EBPO), which regularizes group-relative baselines by “borrowing strength” from accumulated global reward statistics, and in leave-one-out James–Stein baselines that combine prompt-local and across-prompt means (Han et al., 5 Feb 2026, Zeng et al., 5 Nov 2025).

In supervised and reinforcement post-training of reasoning LLMs, the term also denotes a deterioration in coverage. Here the central empirical signature is that pass@1 improves while pass@k degrades, indicating that repeated sampling explores fewer successful reasoning trajectories. The cited literature attributes this behavior to “forks in the road” or decision points at which multiple valid continuations exist but the training data provides only a single demonstrated path, encouraging overconfident early commitment (Nguyen et al., 16 May 2026).

A third usage is compression-oriented. In SmartThinker, reasoning shrinkage means compressing chain-of-thought generation by reallocating tokens from low-importance steps to critical steps. The target is not uniform shortening but differentiated step-level control that preserves correctness-critical computation while reducing redundant verbosity (He et al., 6 Jul 2025).

These usages are connected by a common structural motif: some locally observed signal is unreliable or inefficient, and shrinkage modifies it by pooling information, reweighting alternatives, or imposing selective compression. A plausible implication is that the phrase is best understood as a family resemblance term rather than a single technical primitive.

2. Shrinkage as a stabilization mechanism in RLVR

RLVR optimizes a policy $\pi_\theta$ using verifiable rewards, often binary correctness signals $r \in \{0,1\}$. In critic-free group-relative methods such as GRPO, a prompt $q$ receives $G$ sampled completions, and the standard advantage is

$$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$

with $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$. Policy updates use a PPO-style clipped objective,

$$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$

where $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$. The two failure modes highlighted in EBPO are high variance when $G$ is small and vanishing gradients when all rewards in a group are identical, especially all-zero groups on hard reasoning prompts (Han et al., 5 Feb 2026).

EBPO replaces the purely local baseline with an empirical-Bayes shrinkage estimator,

$$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$

instantiated as

$r \in \{0,1\}$0

Here $r \in \{0,1\}$1 is the within-group reward variance, $r \in \{0,1\}$2 is the variance of prompt means across prompts, and $r \in \{0,1\}$3 is a running estimate of overall success rate. The resulting policy gradient can be written as

$r \in \{0,1\}$4

and in PPO form EBPO simply replaces $r \in \{0,1\}$5 with $r \in \{0,1\}$6 while retaining clipping and optional KL or entropy regularization (Han et al., 5 Feb 2026).

The global statistics are updated online with Welford’s algorithm. For a new observation $r \in \{0,1\}$7, previous count $r \in \{0,1\}$8, mean $r \in \{0,1\}$9, and sum of squared deviations $q$0,

$q$1

$q$2

EBPO applies this once to all rewards to update $q$3 and $q$4, and once to prompt-level group means to update $q$5 (Han et al., 5 Feb 2026).

A parallel line develops an unbiased leave-one-out James–Stein shrinkage baseline for RLVR. For prompt $q$6 and sample $q$7,

$q$8

and the baseline is

$q$9

This construction leaves out both the paired sample and the current prompt, so $G$0 is independent of $G$1 and the policy-gradient estimator remains unbiased. The population-optimal shrinkage intensity has the James–Stein form

$G$2

with $G$3 and $G$4 (Zeng et al., 5 Nov 2025).

The theoretical claims across these RLVR papers are tightly aligned. EBPO proves strictly lower baseline MSE than GRPO when $G$5, bounded entropy decay relative to GRPO, and non-vanishing penalty gradients in saturated failure regimes because all-zero groups produce $G$6 rather than zero. The James–Stein baseline proves lower baseline MSE, hence lower gradient variance under the standard approximation relating gradient variance to baseline MSE, while preserving unbiasedness through leave-one-out independence (Han et al., 5 Feb 2026, Zeng et al., 5 Nov 2025).

3. Empirical consequences in reasoning RL

EBPO is evaluated on DAPO-Math-17K with Pass@1 on AIME-2024/2025, AMC23, MATH-500, and OlympiadBench for LLaMA3.1-8B, Qwen3-8B, and Qwen3-14B. With group size $G$7 and topic clustering, Qwen3-8B average Pass@1 improved from $G$8 under GRPO to $G$9 under EBPO, and across 15 model–dataset combinations EBPO was best in 9. For Qwen3-8B with $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$0, EBPO-topic averaged $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$1 versus $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$2 for GRPO, approximately $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$3 points. The reported advantage is largest for $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$4 at the small-group end, where GRPO’s local-statistics variance is worst (Han et al., 5 Feb 2026).

The same paper reports stability effects that match its theory. EBPO maintains healthy, non-vanishing gradient norms, bounds per-step KL so that late-training spikes observed in GRPO are avoided, and sustains higher policy entropy throughout training. In saturated zero-reward regimes, EBPO generates informative negative advantages proportional to $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$5, whereas GRPO yields zero gradients. Difficulty-stratified curriculum learning further improves generalization on hard reasoning prompts: for Qwen3-8B, EBPO-diff exceeds GRPO-diff by $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$6 on AIME-2024 and $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$7 on AIME-2025. Topic clustering also matters, with EBPO-topic outperforming EBPO-naive on high-difficulty sets, indicating that a coherent prior improves shrinkage quality (Han et al., 5 Feb 2026).

The leave-one-out shrinkage baseline literature reports a similar empirical pattern. On math reasoning with Qwen2.5-Math-1.5B/7B and Qwen3-4B-Base, the JS baseline improves Pass@1 over RLOO by about $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$8 to $$\hat A_i = \frac{r_i - \mu_{\text{group}}}{\sigma_{\text{group}} + \epsilon},$$9 across benchmarks such as MATH500, OlympiadBench, and AMC23. On logic tasks including Knights-and-Knaves, Countdown, and Maze, gains range from $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$0 to $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$1 versus RLOO. On GSM8k with 64 prompts per step and rollout counts $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$2, final test accuracy is reported as JS $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$3 versus RLOO $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$4, GRPO $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$5, REINFORCE++ $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$6, BLOO $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$7, and ReMax $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$8 (Zeng et al., 5 Nov 2025).

Variance-reduction measurements strengthen the estimator interpretation. Relative to RLOO, JS reduces baseline MSE by $\mu_{\text{group}} = \frac{1}{G}\sum_j r_j$9 at $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$0, $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$1 at $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$2, and $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$3 at $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$4 on DAPO17k with Qwen3-4B-Base. Reported gradient-variance reductions are $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$5, $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$6, $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$7, and up to $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$8 across models and datasets. The learned $$J_{\text{RLVR}}(\theta) = \mathbb{E}_{q \sim D}\left[\frac{1}{G}\sum_i \min\big(\rho_i(\theta)\hat A_i,\ \mathrm{clip}(\rho_i(\theta),1-\epsilon,1+\epsilon)\hat A_i\big)\right],$$9 decreases as $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$0 increases and often decays during training as policies become more deterministic, which is consistent with the intended adaptive shrinkage behavior (Zeng et al., 5 Nov 2025).

Operationally, these papers converge on a clear regime of use: small rollout groups, sparse verifiable rewards, and heterogeneous prompt difficulty. In that setting, shrinkage is not used to reduce reasoning diversity but to regularize the control variate that defines the policy gradient.

4. Coverage shrinkage in post-trained reasoning models

A different line of work uses reasoning shrinkage to denote a loss of coverage after post-training. The formal object is pass@k. For correctness indicator $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$1 and decoding distribution $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$2,

$$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$3

and, under the standard independence-of-samples assumption,

$$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$4

Over a dataset $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$5 with $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$6 samples per problem and $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$7 correct solutions for $$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$8, the empirical estimator is

$$\rho_i(\theta)=\pi_\theta(o_i|q)/\pi_{\theta_{\text{old}}}(o_i|q)$$9

Coverage shrinkage is the pattern in which post-training raises pass@1 but lowers pass@k for larger $G$0, showing that the model narrows onto dominant modes while suppressing alternatives (Nguyen et al., 16 May 2026).

The proposed mechanism is data-centric. A decision point or “fork in the road” is a state with multiple valid continuations, only some of which lead to eventual correctness, but where local surface cues do not reliably distinguish them. Under SFT on single-path demonstrations, the model repeatedly encounters indecipherable choices and learns to commit early using spurious cues. This increases pass@1 by concentrating probability on a demonstrated path but decreases pass@k by collapsing diversity (Nguyen et al., 16 May 2026).

The synthetic graph experiments isolate this mechanism. The task uses a star graph with 2 branches and path lengths of 10, with 6,400 training samples and 1,000 test samples, evaluated on Qwen-2.5-0.5B and EvoLM-1B after 16 SFT epochs at learning rate $G$1. In the forward solution style, the first step requires choosing among branches, creating a decision point; in the reverse style, target-to-source reasoning removes the fork. Reverse solutions produce near-perfect pass@k that remains stable across $G$2 and training epochs. Forward solutions instead show steadily improving pass@1 but pass@k that first improves and then degrades, with stronger degradation at larger $G$3. Token-level probing further shows that confidence at the branch-selection token rises sharply for both correct and incorrect branch choices, indicating uncalibrated overconfidence rather than genuine disambiguation. Shuffling variable dependency order while preserving semantics changes the chosen branch and confidence, demonstrating reliance on superficial formatting cues (Nguyen et al., 16 May 2026).

The same paper reports that GRPO applied after one SFT epoch produces the same shrinkage pattern in the forward setting and even more strongly than SFT. This result is used to argue that the driver is not only the optimization algorithm but also the decision-point structure of the training data (Nguyen et al., 16 May 2026).

Real-data experiments exhibit the same geometry. On GSM8K-style math reasoning, the authors compare natural-language and code solution modes using OpenMathInstruct-2 and OpenMathInstruct-1, with identical global $G$4 mode ratios but different organization. When diversity is distributed at the data level, with each problem appearing in a single mode, pass@1 rises monotonically while pass@k degrades sharply over training; when diversity is organized at the problem level, with each problem appearing in both modes, pass@k remains stable with only minor drops at large $G$5. Data-level diversity produces bimodal, overconfident mode commitment per problem, whereas problem-level diversity yields balanced and calibrated mode distributions (Nguyen et al., 16 May 2026).

A related mode-selection phenomenon appears in distilled reasoning models that contain both linear and backtracking traces. On DeepSeek-R1-Distilled variants, small prefix changes such as “Okay,” “Alright,” “Let,” and “To” induce large shifts in reasoning structure, response length, and accuracy. The reported variation reaches up to approximately $G$6 in performance and approximately $G$7 in length solely from different starting tokens. On counterfactual arithmetic, forcing backtracking via prefixes can improve performance by up to $G$8 relative to the default; on factual QA, backtracking harms performance and short linear reasoning works better. The paper interprets these effects as brittle initial decision points controlling reasoning mode (Nguyen et al., 16 May 2026).

Mitigation is correspondingly data- and decoding-centric. Problem-level diversity, multiple valid trajectories per instance, and rationale-aware curation are reported to preserve coverage better than globally diverse but per-instance single-path data. At inference time, uniformly sampling the first reasoning token from the Top-8 high-probability options improves pass@k on later SFT checkpoints and recovers coverage that default decoding loses. The paper’s interpretation is that alternative strategies are not erased but suppressed behind collapsed early decisions (Nguyen et al., 16 May 2026).

5. Compression-oriented shrinkage of chains of thought

SmartThinker introduces a third sense of reasoning shrinkage: compressing reasoning traces while preserving the parts that matter for correctness. Its critique is directed at global length penalties in PPO- or GRPO-style training, which often over-compress critical steps while leaving trivial verbosity intact. The proposed alternative is a two-stage procedure: short-form adaptation via rejection sampling plus SFT, followed by Step-Level Length Control Policy Optimization (SCPO) (He et al., 6 Jul 2025).

In the first stage, the model generates multiple solutions per question, retains the shortest correct response, discards samples longer than 4096 tokens, and trains on an approximately 2,000-example short-CoT corpus using learning rate $G$9, cosine schedule, warm-up ratio $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$0, 3 epochs, batch size 1, and gradient accumulation 8. In the second stage, SCPO uses four components: an online importance estimator, a step-level length control reward, step-level generalized advantage estimation (S-GAE), and difficulty-adaptive clipping (He et al., 6 Jul 2025).

The online importance estimator measures the drop in answer probability when a step is removed. For response $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$1 and step $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$2,

$$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$3

where $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$4 and $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$5. Difficulty is $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$6, and a keyword-aware bonus yields $$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$7. Rewards then weaken penalties on important steps and relax step-count penalties on hard questions. S-GAE computes

$$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$8

and PPO-style clipping is made difficulty-adaptive through

$$b_{\text{EBPO}} = (1-\lambda)b_{\text{Local}} + \lambda \mu_{\text{global}},$$9

The default hyperparameters are $r \in \{0,1\}$00, $r \in \{0,1\}$01, $r \in \{0,1\}$02, and $r \in \{0,1\}$03, with $r \in \{0,1\}$04 so that no KL penalty is used during training (He et al., 6 Jul 2025).

Training uses DeepSeek-R1-Distill-Qwen-1.5B with full fine-tuning and DeepSeek-R1-Distill-Qwen-7B with LoRA, on 3,000 QA pairs from DeepScaleR-Preview. The RL stage uses TRL, ZeRO Stage 2 on two A6000 GPUs, batch size 8, 8 rollouts per question, context length 4K for training and 8K for evaluation, AdamW with $r \in \{0,1\}$05, $r \in \{0,1\}$06, learning rate $r \in \{0,1\}$07, cosine schedule, and a 60-step warm-up. Each batch updates the policy four times (He et al., 6 Jul 2025).

Empirically, the 1.5B model improves average Pass@k from $r \in \{0,1\}$08 to $r \in \{0,1\}$09 and Maj@k from $r \in \{0,1\}$10 to $r \in \{0,1\}$11, while reducing average token usage from 5763 to 3283, or $r \in \{0,1\}$12 of the base length. For the 7B model, average Pass@k rises from $r \in \{0,1\}$13 to $r \in \{0,1\}$14, Maj@k from $r \in \{0,1\}$15 to $r \in \{0,1\}$16, and AvgLen falls from 5159 to 3938, or $r \in \{0,1\}$17 of the base length. The gains are benchmark-dependent: for example, on the 7B model MinervaMATH Pass@5 decreases from $r \in \{0,1\}$18 to $r \in \{0,1\}$19, even as Maj@5 slightly increases from $r \in \{0,1\}$20 to $r \in \{0,1\}$21, illustrating that compression is not uniformly beneficial (He et al., 6 Jul 2025).

The reported distributional changes are also step-specific rather than merely length-specific. The proportion of effective steps increases by $r \in \{0,1\}$22 on low-difficulty problems and $r \in \{0,1\}$23 on high-difficulty problems, while the proportion of total length devoted to effective steps rises by approximately $r \in \{0,1\}$24. SFT warm-up changes training efficiency substantially: without it, 1,000 steps took 122 hours; with it, 1,300 steps took 68 hours. Ablations further indicate that $r \in \{0,1\}$25 yields overly long sequences, overly small $r \in \{0,1\}$26 harms advantage accuracy, $r \in \{0,1\}$27 minimizes tokens but reduces accuracy, and large $r \in \{0,1\}$28 eventually causes collapse (He et al., 6 Jul 2025).

In this usage, shrinkage is explicitly constructive. The model is trained to shrink redundant reasoning while preserving mathematically substantive derivations and essential verification.

6. Statistical lineage and conceptual relations

The shrinkage language in reasoning-model research inherits directly from statistical estimation. One relevant bridge is Prediction-Powered Adaptive Shrinkage (PAS), which treats model predictions as a shrinkage target for multiple mean estimation. For task $r \in \{0,1\}$29, the power-tuned PPI estimator is

$r \in \{0,1\}$30

and PAS then shrinks it toward the prediction mean $r \in \{0,1\}$31 via

$r \in \{0,1\}$32

The global tuning parameter is selected by minimizing the Correlation-Aware Unbiased Risk Estimate (CURE), which is unbiased for compound risk and asymptotically optimal. The paper explicitly frames this as shrinking toward the model’s reasoning, encoded in unlabeled prediction averages (Li et al., 20 Feb 2025).

Bayesian shrinkage theory supplies another background. Global-local priors represent coefficients as

$r \in \{0,1\}$33

with shrinkage factor $r \in \{0,1\}$34. Log-scale priors study $r \in \{0,1\}$35 and show how concentration near $r \in \{0,1\}$36 and tail robustness depend on tail behavior in $r \in \{0,1\}$37-space. The log-$r \in \{0,1\}$38 prior is notable because, for all choices of its scale parameter, the marginal prior diverges at $r \in \{0,1\}$39 and has super-Cauchy tails, and is shown to be KL super-efficient at $r \in \{0,1\}$40 (Schmidt et al., 2018). In a complementary direction, high-dimensional Bayesian shrinkage theory proves that many global-only or exponential-tail global-local priors, including the Bayesian Lasso, are suboptimal for sparse normal means, while Dirichlet–Laplace priors improve prior concentration near sparse vectors and support efficient computation through normalized random measure representations (Bhattacharya et al., 2012).

The same logic extends beyond reasoning models proper. Nonlinear covariance shrinkage and cross-validation-based covariance shrinkage replace noisy sample spectra with bias-corrected estimates that more closely approximate spectral oracles. In the CVC formulation,

$r \in \{0,1\}$41

and isotonic regression is then used to enforce spectral monotonicity. The methodological parallel is that a high-variance local estimate is regularized toward a more stable structure to improve out-of-sample risk (Bartz, 2016).

Taken together, these antecedents clarify why “reasoning shrinkage” can refer to opposed outcomes. In estimator design and step-level compression, shrinkage is a deliberate bias–variance or efficiency trade-off. In coverage studies, shrinkage names an unwanted narrowing of the model’s reasoning support. The shared vocabulary reflects the same mathematical intuition—mass is being concentrated—but the object being concentrated differs: a baseline estimator, a token budget, or a distribution over reasoning paths.