Adaptive Temperature: Dynamic Control

• Adaptive Temperature is the state-dependent adjustment of temperature parameters that modulate distribution sharpness and thermal responses.
• It is applied across machine learning, forecasting, and hardware systems to balance exploration versus concentration and improve calibration.
• Implementations vary from analytic and learned schedules to passive material responses, offering flexible control in diverse application domains.

Searching arXiv for papers on adaptive temperature across machine learning, forecasting, and physical systems. Adaptive temperature denotes the dynamic adjustment of a temperature parameter, or of a temperature-dependent mechanism, so that a model or device changes its behavior with context, uncertainty, training progress, or physical state. In machine learning, the term usually refers to modifying the sharpness of a softmax distribution, an attention map, or a decoding policy; in forecasting, sensing, and thermoregulating materials, it denotes state-dependent estimation, compensation, or heat-transfer control tied to actual thermal conditions (Ohamouddou et al., 27 Aug 2025, Lin et al., 2018, Salami et al., 4 Nov 2025). Across these settings, the common role of temperature is to regulate a trade-off between concentration and dispersion: low temperature sharpens distributions or tightens control, whereas high temperature softens distributions or increases exploratory or dissipative behavior (Lin et al., 2018, Dang et al., 12 Feb 2026).

1. Conceptual and mathematical role

A canonical formulation appears in temperature-scaled softmax. In sequence modeling and reinforcement learning, the sampling distribution can be written as

$$\pi_{\theta,T}(o_t \mid s_t) = \frac{\exp\big(z(o_t \mid s_t)/T\big)}{\sum_{o_t'} \exp\big(z(o_t' \mid s_t)/T\big)},$$

so that low $T$ concentrates mass on high-logit outcomes and high $T$ flattens the distribution (Dang et al., 12 Feb 2026). The same structural role appears in attention mechanisms, where a temperature-scaled attention weight is

$$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$

with low $\tau_t$ yielding concentrated attention and high $\tau_t$ yielding diffuse attention (Lin et al., 2018).

The distinction between fixed and adaptive temperature is central. Fixed temperature assumes that one level of softness is appropriate for all samples, tokens, epochs, or operating conditions. The cited literature repeatedly rejects that assumption. In knowledge distillation, student capacity and training stage alter the desirable sharpness of teacher supervision (Ohamouddou et al., 27 Aug 2025). In calibration, different inputs contribute differently to calibration error and therefore require input-specific confidence scaling (Joy et al., 2022). In code generation and LLM reinforcement learning, different decoding steps and trajectories require different exploration levels (Zhu et al., 2023, Dang et al., 12 Feb 2026). In physical systems, the same principle appears as temperature-responsive emissivity, temperature-dependent hardware compensation, or variable thermal thresholds (Ramirez-Cuevas et al., 2023, Salami et al., 4 Nov 2025, Dowling et al., 2024).

A useful synthesis is that adaptive temperature is not a single method class but a control primitive. It may be epoch-wise, sample-wise, token-wise, trajectory-wise, or embodied in material response. This suggests a unifying interpretation: temperature is a tunable “softness controller” whose optimal setting is rarely stationary.

2. Knowledge distillation and probabilistic calibration

In knowledge distillation, adaptive temperature regulates how strongly a student imitates the teacher’s soft predictions. ATMS-KD defines adaptive temperature as a capacity- and progress-aware schedule that starts high and anneals downward over training, with the initial temperature depending on student width: Compact $T_0=6.0$, Standard $T_0=4.5$, and Enhanced $T_0=4.3$, with empirically observed final temperatures $3.27$, $T$0, and $T$1 after 80 epochs (Ohamouddou et al., 27 Aug 2025). On the Damascena rose maturity classification dataset, the compact student reached $T$2 validation accuracy, $T$3 average inference latency, and $T$4 knowledge retention, outperforming eleven distillation baselines (Ohamouddou et al., 27 Aug 2025). The same paper also makes explicit that adaptive temperature affects only the KD term, not the supervised cross-entropy term, so it changes the character of teacher supervision without changing the hard-label loss.

A different line of work treats adaptive temperature as post-hoc calibration. Sample-dependent adaptive temperature scaling replaces a single global temperature by $T$5, predicted per input, so that easy and hard examples are not uniformly flattened; on ResNet50 and WideResNet28-10 across CIFAR10/100 and Tiny-ImageNet, this improves ECE over global temperature scaling and also yields a hardness signal for datapoints (Joy et al., 2022). The “Adaptive Temperature Scaling” family further generalizes temperature scaling to $T$6, a function of logits, and proposes Entropy-based Temperature Scaling,

$T$7

with strong inductive bias toward entropy-conditioned scaling; the reported result is state-of-the-art performance together with robustness under calibration data scarcity (Balanya et al., 2022).

A conformal variant uses adaptive temperature as a constraint-solving device rather than as a learned regressor. Adaptive Temperature Scaling with Conformal Prediction defines an input-specific temperature $T$8 so that the total probability mass assigned to the conformal prediction set equals $T$9,

$T$0

while preserving the standard conformal marginal coverage guarantee because the temperature changes only the probability allocation, not the set itself (Kotelevskii et al., 21 May 2025).

Another distillation paper derives temperature analytically from logits. “Adaptive Temperature Based on Logits Correlation in Knowledge Distillation” standardizes logits by z-score and sets sample-wise temperature from the maximum standardized logit so that the Taylor-approximation regime holds; under that approximation, the KD objective is governed by teacher–student logits correlation, which gives a direct interpretation of what temperature is controlling (Matsuyama et al., 12 Mar 2025). This suggests that adaptive temperature can be schedule-based, learned, constraint-based, or analytic, with the same underlying purpose: to align the information content of softened distributions with the learning problem.

3. Token-wise control in translation, decoding, and LLM reinforcement learning

In neural machine translation, adaptive temperature appears as an attention-control mechanism. The Self-Adaptive Control of Temperature model computes a decoding-step-specific attention temperature

$T$1

so that $T$2 and attention can become sharply concentrated for content words or diffuse for function words (Lin et al., 2018). The reported gains are substantial: on Chinese–English, Seq2Seq + SACT improves the authors’ Seq2Seq baseline from $T$3 to $T$4 BLEU, and on English–Vietnamese from $T$5 to $T$6 BLEU (Lin et al., 2018). The accompanying qualitative analysis shows high temperatures for tokens such as “to”, “from”, “they”, and punctuation, and low temperatures for lexical items such as “pay attention” and “nuclear,” grounding the mechanism in word-type-sensitive alignment behavior.

For code generation, the temperature adaptation target shifts from attention to sampling. “Hot or Cold? Adaptive Temperature Sampling for Code Generation with LLMs” reports that token-level loss in code is highly non-uniform: challenging tokens cluster at the beginning of code blocks, whereas many subsequent tokens are comparatively deterministic (Zhu et al., 2023). AdapT therefore uses a larger temperature for code block initial tokens and a smaller temperature elsewhere. On HumanEval, CodeGeeX-13B improves from $T$7 to $T$8 pass@15, and on MBPP InCoder-6B improves from $T$9 to $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$0 pass@15 (Zhu et al., 2023). The paper also reports reductions in SyntaxError, TypeError, and NameError, which supports the interpretation that low temperature suppresses tail randomness on tokens that should be nearly deterministic.

In LLM post-hoc calibration after RLHF, adaptive temperature becomes token-specific confidence correction. “Calibrating LLMs with Adaptive Temperature Scaling” predicts a temperature for each token from last-layer hidden states and fits the calibration head on a standard SFT dataset; the reported effect is a calibration improvement of over $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$1–$$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$2 across three downstream benchmarks without impeding performance improvements from RLHF (Xie et al., 2024). This contrasts with global temperature scaling, which is too coarse for the heterogeneous calibration shift induced by RLHF.

In LLM reinforcement learning, temperature itself becomes a meta-policy. Temperature Adaptive Meta Policy Optimization defines a distribution over candidate temperatures $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$3, updates the base policy in an inner loop, and updates the temperature meta-policy in an outer loop using temperature-specific trajectory advantages derived from likelihoods and rewards (Dang et al., 12 Feb 2026). On five mathematical reasoning benchmarks, the average Pass@1/Pass@8 of the base model is $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$4, the best fixed or heuristic GRPO baselines reach $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$5 or $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$6, and TAMPO reaches $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$7 (Dang et al., 12 Feb 2026). Here adaptive temperature is no longer merely a softmax modifier; it is an explicit exploration controller trained from trajectory evidence.

4. Forecasting, thermometry, and adaptive estimation

Outside representation learning and decoding, adaptive temperature often refers to modeling or measuring a thermal variable under nonstationary conditions. In sea-surface-temperature forecasting, the PCA-enhanced Adaptive NVAR framework compresses SST fields with SVD and replaces the fixed quadratic nonlinear feature map of standard NVAR with a learned MLP feature map $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$8, while preserving delay embedding and linear readout structure (Azimov et al., 10 Jun 2026). The reported result is that Adaptive NVAR consistently achieves lower forecasting errors across multiple prediction horizons, and SVD reduces computational complexity, producing a fast and scalable framework suitable for real-time forecasting (Azimov et al., 10 Jun 2026). In this setting, “adaptive temperature” means adaptive modeling of temperature dynamics rather than adaptive softmax scaling.

In ensemble weather forecasting, spatial adaptivity addresses local thermal structure rather than time-varying softmax behavior. “Spatially adaptive post-processing of ensemble forecasts for temperature” extends non-homogeneous Gaussian regression by expressing the predictive mean as local short-term temperature averages plus ensemble-driven anomalies, and by interpolating local averages and local uncertainty parameters with intrinsic Gaussian random fields that include location-dependent nugget effects (Scheuerer et al., 2013). The reported outcome is locally calibrated and sharp probabilistic forecasts that compare favorably with alternatives such as geostatistical model averaging and remain computationally efficient for operational use (Scheuerer et al., 2013). A plausible implication is that the “adaptive” qualifier can refer to spatial localization of statistical structure rather than explicit time-varying control.

Adaptive Bayesian thermometry makes the control variable experimental rather than algorithmic. In release–recapture thermometry of few ultracold $$\tilde{\alpha}_{t,i} = \frac{\exp(\tau_t^{-1} e_{t,i})}{\sum_{j=1}^{n} \exp(\tau_t^{-1} e_{t,j})},$$9 atoms, the temperature estimate is updated with a Bayesian posterior and the next release–recapture time is chosen to maximize expected information gain, thereby avoiding uninformative measurements (Glatthard et al., 2022). The authors report that adaptively choosing the release–recapture times substantially reduces the number of measurements needed for convergence, and detailed simulations show that an a priori optimized Bayesian protocol reduces variability by about $\tau_t$0 relative to an unoptimized Bayesian protocol and by more than $\tau_t$1 relative to the conventional fit-based method in one studied setting (Glatthard et al., 2022). Here the adaptive object is not the temperature parameter in a softmax, but the experimental policy used to estimate a physical temperature efficiently.

5. Thermoregulation, sensing hardware, and thermal control systems

In sensing hardware, adaptive temperature may denote internal thermal compensation. “Adaptive Internal Calibration for Temperature-Robust mmWave FMCW Radars” models the dependence of the amplitude profile on internal radar temperature with per-antenna, per-bin linear regression and applies multiplicative correction to obtain Temperature Compensated Amplitude Profiles (Salami et al., 4 Nov 2025). Across two settings and three antennas, Pearson correlations between internal temperature and raw amplitude profile are approximately $\tau_t$2–$\tau_t$3, and the proposed scheme reduces the correlation by up to $\tau_t$4 (Salami et al., 4 Nov 2025). The method is explicitly internal-only: it uses the radar’s own IF measurements and internal temperature sensor, not external references.

In passively thermoregulating materials, adaptive temperature is embodied directly in material response. The passively adaptive radiative switch for buildings uses hexadecane-driven louver actuation to switch between a black heating state and a white radiative-cooling state within less than $\tau_t$5, and an outdoor test reports a $\tau_t$6 reduction in cooling energy cost and a $\tau_t$7 reduction in heating energy cost relative to non-switching devices (Xiao et al., 2023). The infrared thermochromic antenna composite based on VO$\tau_t$8 rods, stars, and flakes reports a massive $\tau_t$9-fold increase in absorption cross-section for non-spherical antennas as temperature rises, and composite emissivity switching of $\tau_t$0 in tunable spectral ranges (Ramirez-Cuevas et al., 2023). In both cases, the adaptive mechanism is a passive feedback between temperature and radiative coupling.

A more actively controlled material system appears in deep learning-assisted thermal metamaterials. There, an infrared camera measures ambient temperatures, a neural network outputs the target angular velocity $\tau_t$1, and a rotating PDMS core changes effective thermal conductivity $\tau_t$2, enabling spontaneous response over wide working temperatures and robustness to changes in the external thermal-field direction (Jin et al., 2023). This suggests a different design pattern: thermal adaptivity mediated by learned control rather than by phase transition alone.

In computing hardware, adaptive temperature can be a scheduling variable. VTF-TAS introduces a variable temperature threshold $\tau_t$3 for thermal-aware scheduling, updated from a heuristic derived from fluid scheduling. The reported result is that with $\tau_t$4, peak temperature is $\tau_t$5, about $\tau_t$6 lower than POD-TAS for the same workload, without requiring the expensive search used by POD-TAS (Dowling et al., 2024). In quantum thermodynamics, adaptivity is structural: an adaptive quantum heat engine couples a three-level working medium to a controller that shifts level spacings so the engine can always operate for variable bath temperatures, and the paper also derives conditions for maximum power extraction for all bath temperatures (Khanahmadi et al., 2021). Across these examples, adaptive temperature may be a calibration law, a phase-transition response, a controller state, or a scheduling threshold, but in each case it serves to preserve functionality under thermal variability.

6. Recurrent design patterns, limitations, and common misconceptions

A recurrent pattern is that adaptive temperature is almost never just “a better fixed temperature.” The surveyed work includes hand-designed curricula over epochs (Ohamouddou et al., 27 Aug 2025), per-sample post-hoc scaling (Joy et al., 2022), entropy-conditioned two-parameter calibration (Balanya et al., 2022), per-token attention control (Lin et al., 2018), block-aware decoding schedules (Zhu et al., 2023), trajectory-level meta-policies (Dang et al., 12 Feb 2026), analytic temperature rules derived from logits geometry (Matsuyama et al., 12 Mar 2025), and passive temperature responses in materials (Xiao et al., 2023, Ramirez-Cuevas et al., 2023). A common misconception is therefore that adaptive temperature implies a single implementation template. The literature instead presents a family of control strategies whose only invariant is that temperature is made state-dependent.

A second recurring theme is that adaptivity trades expressiveness against robustness. In calibration, expressive neural temperature functions outperform simpler methods when calibration data are abundant, but they can fail under data scarcity, whereas entropy-based or otherwise strongly biased parameterizations remain robust (Balanya et al., 2022). In distillation, hand-designed adaptive schedules improve performance but remain task- and architecture-specific; ATMS-KD explicitly notes task-scope limitations, hand-crafted annealing, and possible retuning needs for multi-class or multi-task settings (Ohamouddou et al., 27 Aug 2025). In sequence models, SACT is demonstrated on RNN-based Seq2Seq and notes that transfer to Transformer multi-head attention requires additional design choices, while TAMPO shows that greedy temperature selection harms exploration and that the candidate temperature range must remain reasonable (Lin et al., 2018, Dang et al., 12 Feb 2026).

A third pattern is that adaptivity does not eliminate domain constraints. The radar calibration paper focuses on amplitude drift rather than the full phase/frequency problem (Salami et al., 4 Nov 2025). The radiative switch for buildings notes durability concerns for moving parts and environmental sensitivity (Xiao et al., 2023). The VO$\tau_t$7 thermochromic composites switch near $\tau_t$8, which is advantageous for some applications but implies application-specific transition engineering for others (Ramirez-Cuevas et al., 2023). The cold-atom thermometry work shows that a simpler non-adaptive protocol exploiting all prior information can be competitive in some regimes, so full adaptivity is not always necessary (Glatthard et al., 2022).

Taken together, these results suggest that adaptive temperature is best understood as a control-theoretic device for matching softness, confidence, exploration, or heat-transfer response to local conditions. In learning systems, it governs the information geometry of distributions; in physical systems, it governs how structure or response changes with thermal state. The unifying claim supported by the cited work is not that one adaptive-temperature mechanism dominates all others, but that state-dependent temperature control repeatedly yields measurable gains whenever fixed-temperature assumptions are too coarse for the underlying heterogeneity.