Token Entropy Cumulative Average (TECA) Overview

• Token Entropy Cumulative Average (TECA) is a metric that computes the running average of token-level uncertainty during language model generation.
• It provides insights into reasoning dynamics by capturing the shift from initial high uncertainty in exploration to stabilized, decisive token predictions.
• TECA is applied in reinforcement learning and overthinking mitigation strategies, enabling efficient termination and refined credit assignment in generative models.

Token Entropy Cumulative Average (TECA) is a sequence-level uncertainty statistic for language-model generation defined as the running average of token-level predictive entropy. In the formulation used for reasoning LLMs, token entropy at step $t$ is $H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$, and TECA is $\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$, where $V$ is vocabulary size and $p_{t,j}$ is the model probability for token $j$ at step $t$. The metric is used to characterize the cumulative exploration state of a generation trace, especially in chain-of-thought reasoning, and closely relates to sequence-level average entropy formulations that appear under other names such as “Mean EAS” and “sequence entropy” (Jiang et al., 2 Oct 2025, Zhu et al., 28 Aug 2025).

1. Definition and formalization

TECA is built from token-level predictive entropy. For a model generating the $t$-th token, the token entropy is

$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}.$

The cumulative average up to position $t$ is then

$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$0

This quantity is computed online during inference: at each generation step, the model entropy $H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$1 is evaluated from the next-token distribution, the cumulative sum is updated, and the result is divided by the current step index. In this form, TECA measures the average uncertainty, or average exploration, in the sequence so far (Jiang et al., 2 Oct 2025).

Equivalent or closely related formulations appear in adjacent work. In token- and sequence-level RL reward shaping, the arithmetic mean sequence entropy is defined as

$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$2

with an optional geometric mean variant, although the arithmetic mean is the one mainly used. In uncertainty modeling for reasoning LLMs, “Mean EAS” is defined as

$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$3

and is described as corresponding directly to what is commonly referred to as TECA (Tan et al., 6 Aug 2025, Zhu et al., 28 Aug 2025).

A normalized entropy variant also appears in entropy-adaptive RL. There, token entropy is divided by $H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$4 to obtain

$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$5

making the entropy invariant to vocabulary size. This normalization is used for discounting and horizon control rather than as the canonical TECA definition, but it preserves the same underlying notion of cumulative token uncertainty (Chawla et al., 3 Jun 2026).

2. Interpretation in reasoning dynamics

TECA is used as a process-level signal for distinguishing exploratory from determined reasoning. In the “Explore Briefly, Then Decide” formulation, early reasoning is described as an exploration stage in which models exhibit high uncertainty, high entropy, and rising TECA, whereas later reasoning enters a determination stage in which token entropy and TECA flatten or drop once the reasoning path has been established (Jiang et al., 2 Oct 2025).

Within this interpretation, overthinking is linked to prolonged, unnecessary increases in TECA. The central claim is that unrestricted chain-of-thought models may continue exploring after the answer has effectively been found, producing long, redundant continuations. By contrast, shorter-thinking models exhibit TECA profiles that rise and then fall or plateau, indicating brief exploration followed by commitment. This yields a directly observable distinction between exploratory indecision and decisive termination (Jiang et al., 2 Oct 2025).

A closely related interpretation appears in sequence-level uncertainty modeling. There, integrating token entropy across the generation path is said to capture both the intensity and duration of uncertainty. High aggregate entropy indicates prolonged or repeatedly high uncertainty, whereas low aggregate entropy indicates early, sustained confidence and stability. Although that work emphasizes the cumulative sum rather than the cumulative average, it supports the broader reading of TECA-like measures as summaries of how uncertainty evolves through a reasoning trace (Zhu et al., 28 Aug 2025).

This suggests that TECA is best understood not as a confidence score for an isolated token, but as a trajectory statistic: it summarizes how much uncertainty has been sustained over the prefix generated so far. In reasoning tasks, that trajectory-level view is precisely what makes TECA useful for identifying when a model should continue exploring and when it should terminate.

3. Relation to adjacent entropy aggregations

TECA belongs to a family of token-entropy aggregations that differ mainly in whether they average, sum, or take an extremum over token uncertainties. Related work makes these distinctions explicit (Zhu et al., 28 Aug 2025, Cheong et al., 19 Mar 2026).

Aggregation	Formula	Role in cited work
TECA / Mean EAS	$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$6	Average per-token uncertainty
EAS	$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$7	Cumulative uncertainty
EntropyCache trigger	$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$8	Per-step recompute signal

The most direct comparison is between TECA and Entropy Area Score (EAS). EAS is the area under the entropy curve,

$H_t := -\sum_{j=1}^{V} p_{t,j} \log p_{t,j}$9

while Mean EAS, identified with TECA, is the per-token average. Their relationship is exact:

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$0

The distinction is interpretive as well as algebraic. EAS measures total cumulative uncertainty, whereas TECA measures average uncertainty per token. The EAS paper states that the average can miss temporally localized uncertainty: if a model is confident for long stretches and uncertain for a short but critical period, the average may mask that behavior, while the sum still accumulates the uncertainty spike (Zhu et al., 28 Aug 2025).

A separate contrast appears in diffusion LLM caching. EntropyCache does not use cumulative or average entropy in its skip-or-recompute rule. Instead, it uses only the per-step maximum entropy of newly decoded token distributions,

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$1

and explicitly states that no cumulative or average entropy such as TECA is used in the main decision mechanism. This provides a useful boundary case: token entropy averages are not universally optimal, and some systems instead require an instantaneous, worst-case signal (Cheong et al., 19 Mar 2026).

4. TECA in reinforcement learning and post-training

Average token entropy has been incorporated into several RL formulations for LLMs, sometimes explicitly as TECA and sometimes under closely related sequence-level averages. In GTPO and GRPO-S, token entropy is used for reward shaping under the premise that high-entropy tokens in correct responses can guide the policy toward a higher performance ceiling. The sequence-level version defines average entropy as

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$2

and uses it in the entropy-weighted reward

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$3

The token-level version instead redistributes reward to individual positions via entropy-weighted token rewards. In both cases, entropy averages function as proxies for uncertainty and exploration, enabling finer-grained credit assignment than uniform sequence rewards (Tan et al., 6 Aug 2025).

In Adaptive-Horizon GRPO, the operative quantity is the cumulative sum of normalized token entropies, not the cumulative average itself. The per-step discount is

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$4

and the cumulative weight is

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$5

That work nevertheless makes the connection to TECA explicit by defining

$\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$6

and stating that the cumulative sum used in the discount is essentially TECA multiplied by prefix length. The practical interpretation is that high TECA implies a trajectory has been generally uncertain so far, which in turn causes stronger horizon shrinkage and down-weighting of later-token gradients. The selective variant, SA-AH-GRPO, applies this discount only to negative-advantage rollouts and reports on the 3B model a peak Pass@1 of $\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$7 at step 30, maintenance of $\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$8 at 180 steps, and training variance reduced to $\mathrm{TECA}_t := \frac{\sum_{k=1}^{t} H_k}{t}$9, described as a $V$0 times reduction relative to GRPO while matching its peak accuracy (Chawla et al., 3 Jun 2026).

STARE uses a batch-level average entropy,

$V$1

described as “Token Entropy Cumulative Average, TECA-style,” as the feedback variable in a closed-loop entropy controller. Reweighting is activated when $V$2, and the stated objective is to stabilize policy entropy within a target band during RL training. In this setting, TECA-style averaging serves as a control signal for entropy preservation rather than a measure of reasoning length or answer uncertainty (Luo et al., 17 Jun 2026).

5. TECA and overthinking mitigation

The most explicit TECA-centered application is the mitigation of overthinking in chain-of-thought reasoning. The stated problem is that LLMs often generate unnecessarily lengthy reasoning steps for simpler problems, which increases computational inefficiency, can reduce accuracy by rethinking after finding a correct answer, and fails to adapt reasoning depth to problem complexity. TECA is introduced as a process-level metric to characterize exploration versus determination more effectively than hard limits on step or token count (Jiang et al., 2 Oct 2025).

This application couples TECA to a reinforcement-learning mechanism called Cumulative Entropy Regulation (CER). The reward combines answer correctness with a TECA-based term:

$V$3

with $V$4 if the answer is correct and $V$5 otherwise, and

$V$6

where $V$7 is the last-step TECA in the output sequence. The effect is segmented: correctness is primary, but among correct answers, lower average exploration yields higher reward (Jiang et al., 2 Oct 2025).

Empirically, the cited results report substantial reductions in response length on mathematical benchmarks while retaining or improving performance on some settings. The paper states that the average response length decreases by up to $V$8 on simpler datasets. The summary table includes the following examples:

Model and dataset	Original length	CER length
Qwen3-4B, GSM8K	1348.59	391.08
Qwen3-8B, GSM8K	1491.38	668.06

The corresponding length reductions are reported as $V$9 and $p_{t,j}$0, with accuracies of $p_{t,j}$1 to $p_{t,j}$2 for Qwen3-4B on GSM8K and $p_{t,j}$3 to $p_{t,j}$4 for Qwen3-8B on GSM8K (Jiang et al., 2 Oct 2025).

Qualitatively, CER-trained models are described as making decisive conclusions shortly after discovering the correct answer, often after 1–2 reflections rather than 5–6. On simple questions, such models may skip unnecessary diagnostic steps entirely. In TECA terms, successful regulation produces a curve that rises during initial exploration and then falls, matching the intended “explore briefly, then decide” pattern (Jiang et al., 2 Oct 2025).

6. Scope, limitations, and terminological ambiguities

TECA is not a single universally adopted label. Closely related metrics appear as “Mean EAS,” “sequence entropy,” “batch-averaged entropy,” and “TECA-style” average entropy. In at least one case, the paper explicitly states that it does not discuss a metric named TECA verbatim, even though it introduces Mean EAS and says that this corresponds directly to what is commonly referred to as Token Entropy Cumulative Average (Zhu et al., 28 Aug 2025).

A substantive limitation of average entropy aggregation is that it can be less sensitive than cumulative entropy to brief but critical uncertainty spikes. This is the central reason the EAS paper argues that total entropy can be more discriminative for answer entropy correlation and data selection. The same paper also states that EAS is best suited for tasks with unique or structured answers such as math, MCQ, and some science QA, and that for free-form text or code, where correctness is semantic rather than token-matching, correlation with output variability drops. A plausible implication is that TECA, as the average counterpart to EAS, is likewise most informative when token-level uncertainty aligns closely with task correctness (Zhu et al., 28 Aug 2025).

Another common confusion is acronymic rather than conceptual. “TECA” also denotes “Test-time Enhancer and Classifier Adaptation” in computer vision, a method that combines an image enhancement model with a classifier and applies uncertainty-aware logit switching during test-time adaptation. That usage is unrelated to Token Entropy Cumulative Average, despite the identical acronym (Enomoto et al., 2024).

Finally, TECA should not be treated as a default uncertainty signal in every generation system. EntropyCache explicitly rejects cumulative and average entropy for its core decision rule, using only the maximum entropy of newly decoded token distributions at each denoising step. This underscores a broader methodological point: TECA is most natural when the objective depends on sustained reasoning dynamics, trajectory-level uncertainty, or cumulative exploration, rather than on instantaneous cache drift or stepwise control (Cheong et al., 19 Mar 2026).