Dynamic Outlier Truncation (DOT)

• Dynamic Outlier Truncation (DOT) is a statistical method that adaptively identifies and suppresses outliers using online estimates and a three-sigma rule.
• It employs a sliding window approach to adjust time-varying control limits for real-time anomaly detection and truncates redundant tokens in RL-finetuned models.
• DOT has demonstrated significant token savings and improved pass@1 accuracy in both microgrid operations and large-scale language model training while maintaining low computational overhead.

Dynamic Outlier Truncation (DOT) is a statistical intervention technique with distinct instantiations for online anomaly detection in streaming systems and for improving efficiency in reinforcement-learning-based reasoning models. In both contexts, it targets process outliers in a principled, dynamically adaptive manner—either by establishing time-varying control limits in real-time environments or by truncating the extreme tails of token-length distributions for already-solved prompts in RL-finetuned LLMs. DOT leverages online statistical estimates and extreme-value thresholding (typically a three-sigma rule) to mitigate overextension—whether by suppressing unwarranted anomalies or by reducing unnecessary verbosity. The method’s statistical guarantees, computational efficiency, and minimal domain-specific assumptions underpin its applicability to both engineering systems and large-scale model training (Wadinger et al., 2023, Wu et al., 7 Jan 2026).

1. Underlying Principles and Formal Definitions

Streaming Anomaly Detection Context. DOT utilizes an online inverse cumulative distribution function (ICDF) approach under a sliding window model. For a real-valued data stream $x_1, x_2, \dots$, at each time $t$ it maintains estimates of the empirical mean $\mu_t$ and variance $\sigma_t^2$ using a window of the most recent $t_e$ samples. Gaussian modeling enables closed-form expressions for the cumulative distribution function (CDF) $F_{X_t}(x)$ and its inverse (percent-point function, PPF). For a target quantile $q$ (commonly aligned with three-sigma for $\alpha \approx 0.0027$), the dynamic limits are

$$U_t = \mu_t + \sigma_t z_{q}, \quad L_t = \mu_t - \sigma_t z_{q}, \quad z_q = \Phi^{-1}(q)$$

where $\Phi^{-1}$ denotes the standard normal PPF. These process limits enable pointwise outlier detection with computational complexity $O(1)$ per arrival (Wadinger et al., 2023).

Reinforcement Learning for Efficient Reasoning Context. In RL-finetuned generation, DOT is applied during training to rollout groups produced via Group-Relative Policy Optimization (GRPO). When a batch of $G$ rollouts $\{o_i\}_{i=1}^G$ for prompt $q$ is found to be "all-correct," DOT computes the empirical mean $\mu_L$ and standard deviation $\sigma_L$ of the rollout lengths $L_i = |o_i|$ and sets a cutoff

$T(q) = \lfloor \mu_L + \alpha \sigma_L \rfloor$

with $\alpha \approx 3$ and margin $m$ to prevent trivial truncations. Only rollouts in all-correct groups with $L_i - T(q) \geq m$ are truncated to $T(q)$ tokens. Rewards are recomputed post-hoc, and standard policy updates proceed unchanged (Wu et al., 7 Jan 2026). This intervention suppresses the extreme tail of redundant tokens while conserving exploratory capacity for prompts where correctness remains unsolved.

2. Statistical Mechanics and Online Estimation

In the streaming context, online Welford recurrences drive numerically stable updates for mean and variance over a sliding window: $$\mu_t = \mu_{t-1} + \frac{x_t - \mu_{t-1}}{n_t}, \quad S_t = S_{t-1} + (x_t - \mu_{t-1})(x_t - \mu_t), \quad \sigma_t^2 = \frac{S_t}{n_t-1}$$ where $n_t \leq t_e$ and the oldest datum is removed with an inverse Welford step when the window is full (Wadinger et al., 2023). This guarantees rapid adaptation to nonstationarity, with $t_e$ (expiration period) controlling memory of past behavior.

In the RL application, DOT truncation operates only on the top tail (rare, large outlier groups), affecting less than $0.5\%$ of rollouts but leading to system-wide entropy collapse and token budget savings. Hyperparameters $\alpha$ and $m$ respectively control the extremity of x-length cutoff and the minimum truncation enforced (Wu et al., 7 Jan 2026).

3. Auxiliary Mechanisms: Drift Adaptation and Training Stabilization

Change-Point and Drift Adaptation (Streaming). The method includes an auxiliary window (time constant $t_c$) over the last $t_c$ anomaly scores $y_t$ to monitor persistent outlier occurrence. If the mean anomaly score in this window rises above the tail quantile $q$, DOT adaptively updates its statistics even in the presence of detected anomalies, facilitating rapid convergence past change-points and enabling ongoing drift adaptation (Wadinger et al., 2023).

KL Regularization and Predictive Sampling (RL). To counteract system-wide entropy collapse, DOT incorporates targeted KL regularization (KL-Cov), penalizing determinization of high-advantage tokens: $$\mathcal{L}_{\mathrm{KL}} = \lambda\,\mathbb{D}_{\mathrm{KL}}(\pi_\theta \| \pi_{\theta_{\mathrm{old}}})$$ with $\lambda \approx 2 \times 10^{-3}$, selectively applied to token sets with strong log-probability/advantage covariance. Predictive dynamic sampling maintains effective batch diversity by oversampling prompts proportionally to the sliding-window mean $\bar p$ and variance $s_p$ of post-DOT "effective group" ratios, with

$$\gamma = \max\left(1, \frac{1}{\bar p(1 + s_p)}\right)$$

and batch assembly strictly of effective, nontrivial groups (Wu et al., 7 Jan 2026).

4. Empirical Results and Case Studies

The empirical efficacy of DOT has been demonstrated in both microgrid operations and reasoning model training:

Streaming Case Studies (Wadinger et al., 2023)

• Battery Energy Storage System (BESS):
  • Data: 1-min sampling, $t_e = 7$ days, $t_c = 5$ hours, $q = 0.9973$.
  • Real-time latency: $\approx$0.1 ms/sample (Python, M1, ICDF via Brent).
  • All eight known anomalies (packet dropouts, sensor faults, BESS relocation, peak tests) were flagged with no false positives.
  • Process limits adapt within one day after change-point (March 7 relocation).
• Power Inverter Temperature:
  • Period: March 16–April 17, 2022, same hyperparameters as BESS.
  • Real-time latency: $\approx$0.08 ms/sample.
  • Correctly flagged four sensor faults and rare heating events after packet loss. The $t_c$ window prevented over-relaxation of bounds from single-point spikes.

RL Reasoning Model Results (Wu et al., 7 Jan 2026)

• On Qwen-1.5B, AIME-24: DOT reduces average length from 15,498 to 5,151 tokens (66% saving), simultaneously increasing pass@1 accuracy from 30.0% to 52.2%.
• On Qwen-7B, DOT-8K achieved 62.6% pass@1 at 4,903 tokens (37% of original length).
• On 32B models, DOT-8K set a new state of the art at 73.2% on AIME-24 (4,151 tokens, ≈40% of baseline).
• Code completion tasks (HumanEval, LiveCodeBench): ≈50% token savings and +4 percentage point pass@1 increase.
• DOT produces system-wide truncation of verbosity, affecting less than 0.5% of rollouts, while preserving long-horizon reasoning.

5. Algorithmic Workflow and Implementation

• Streaming Anomaly Detection (Pseudocode excerpt):

1. For each sample $x_t$, compute anomaly score $y_t$ and update the rolling buffer and scores.
2. If $y_t \geq q$, report outlier; else report normal.
3. Adapt distributional estimates if $y_t < q$ or mean of $y_{t-t_c+1:t}$ exceeds $q$ (to adapt past drift/change-point).
4. Compute $L_t$, $U_t$ as dynamic process bounds, feeding directly to existing PLC/SCADA alerting.

• RL Truncation Intervention:

1. For each batch prompt $q$ with rollout group $\{o_i\}, \{R_i\}$, check for "all-correct" ($R_i=1$).
2. If so, compute group mean and stdev of lengths, and cutoff $T(q)$.
3. For each $o_i$: if $|o_i| - T(q) \geq m$, truncate to $T(q)$; otherwise, retain.
4. Recompute rewards post-truncation. Select effective groups, monitor group-variance ratio, and dynamically adjust batch size.
5. Update the policy with the GRPO objective and KL-Cov regularization (Wu et al., 7 Jan 2026).

6. Deployment Characteristics, Limitations, and Future Directions

DOT is notable for its parameter simplicity: only the sliding window length $t_e$, time constant $t_c$, and tail-quantile $q$ are required in the streaming context, all of which are operationally interpretable. In RL, $\alpha$ (tail width) and $m$ (minimum truncation margin) control intervention aggressiveness; $\alpha = 3$ is robust for most settings. No offline training, domain-specific thresholds, or expensive computation is required. DOT is readily embedded in edge devices, PLCs, or high-throughput training code.

A recognized limitation is dependency on initial policy quality and reward verifier fidelity in RL; public models with low entropy may provide limited headroom for DOT gains. Extension of DOT beyond token-level spaces to agentic tool-calling remains unexplored but is a plausible area for future inquiry. In both contexts, DOT functions as a statistically grounded, minimal yet systemically effective outlier suppression mechanism, avoiding indiscriminate penalization while promoting operational efficiency (Wadinger et al., 2023, Wu et al., 7 Jan 2026).