Agentic Control in Variational Language Models

Yves Ruffenach

cs.LGApr 14, 2026

Agentic Control in Variational Language Models

Authors: Yves Ruffenach

Organizations: Conservatoire National des Arts et Métiers, Strasbourg, France

Abstract

We study whether a variational language model can support a minimal and measurable form of agentic control grounded in its own internal evidence. Our model combines local variational hidden computation (EVE), a homeostatic latent regulator, structurally aware checkpoint retention and a calibrated uncertainty-aware controller operating on top of the retained model. Rather than treating uncertainty as a passive diagnostic measured after prediction, we treat it as an operational signal that can regulate training, support checkpoint retention and guide inference-time intervention. The resulting framework is deliberately focused. It studies a closed-loop form of internal control in which structural and predictive signals become actionable. Empirically, the variational backbone improves over a matched deterministic reference on the language-modeling task while also exhibiting a richer and more usable uncertainty profile. On top of this backbone, the calibrated controller remains active, uses multiple actions under a full agentic evaluation and yields a positive quality-cost trade-off. These results support a precise claim: internal uncertainty can serve not only as a descriptive property of a variational language model, but also as a practical control interface for regulation, checkpoint retention and minimal agentic routing.

Figures & tables

Figure 1: A minimal agentic loop. Internal evidence is observed, mapped to an action, and followed by stronger intervention when required. In the full method, this loop is embedded in a broader pipeline that also includes structurally aware checkpoint retention.

Figure 2: A building metaphor for underused and recruited depth. Before control, the model learns but deeper floors remain dark. After closed-loop intervention, a deeper floor becomes active.

Figure 3: A thermostat metaphor for homeostatic latent regulation. The autopilot keeps the latent regime in a healthy zone: active, stable, and controlled.

Figure 4: Before/after view of the observed system behavior. The variational backbone supports prediction and exposes internal uncertainty signals that are later used by the calibrated controller.

Model	Loss	CE	PPL	Acc.	KL	Local
EVE	5.5340	5.4341	229.09	0.1632	282.2273	0.1309
DET	5.6035	5.6035	271.36	0.1560	0.0000	0.0000

Table 1: Final validation comparison between EVE and DET on the new run.

Figure 5: Observed backbone differences between EVE and DET on selected validation metrics. Positive bars indicate an improvement of EVE relative to DET .

Seed	Best epoch	Val CE	Val PPL	Val Acc.	Val KL	Val Local
EVE S101	3	5.4241	226.82	0.1657	255.1250	0.1338
EVE S202	3	5.4161	225.00	0.1654	304.1061	0.1404

Table 2: Final EVE seed comparison on the new run.

Figure 6: Observed training and validation cross-entropy trajectories for the retained run.

Candidate	$\lambda_{\text{band,high}}$	CE	Local	Frac. High	$\mu^{2}$
Stage1 01	2.00	5.4111	0.1318	0.0856	0.0497
Stage1 02	2.05	5.4094	0.1264	0.0821	0.0484
Stage1 03	2.10	5.4381	0.1267	0.0855	0.0498

Table 3: Stage-1 EVE search summary on the new run.

Model	NLL	ECE	MI	Epi.	Flip	CVaR-NLL	Frac. High	$\mu^{2}$
EVE	5.6404	0.0326	0.1777	0.0325	0.3197	12.5035	0.0826	0.0487
DET	5.6698	0.0792	0.0000	0.0000	0.0000	13.0499	0.0000	0.7370

Table 4: Extended uncertainty and diversity summary on the new run.

Figure 7: Epoch-level latent-state diagnostics for the retained EVE checkpoint family. The figure tracks local activity, the fraction of high-activity units, and latent energy across epochs.

Metric	Value
Selected seed	101
Number of examples	200
Accepted examples	180
Abstained examples	20
Coverage	0.9000
Accepted accuracy	0.1778

Table 5: Full agentic evaluation summary on the new run.

Figure 8: Observed behavior of the calibrated controller in the full multi-action evaluation. The summary box reports coverage, accepted/total examples, and avoided errors relative to direct prediction.

Criterion	Passed	Value
Coverage is sufficient	True	0.9000
Retrieve is used	True	0.3200
Deliberate is used	True	0.2650
Not abstain-dominated	True	0.1000
Utility is positive	True	0.1048
Avoids some direct errors	True	0.0994

Table 6: Agentic verification results on the new run.

Explore similar work

Jun 30, 2026

Harnessing the Latent Space: From Steering Vectors to Model Calibrators for Control and Trust

Language models have changed from unreliable text generators to highly-capable large models with trillions of parameters. Capability increases come hand-in-hand with increases in scale, making understanding the internal representations of models more challenging. Since millions of users increasing rely on language models to interact with external tools or make decisions in medium or high-stakes scenarios, we need to establish control over model behavior and know when to trust model outputs. In this paper, we discuss our contributions on harnessing the latent spaces by proposing steering vectors for control and developing latent space-based model calibrators for trust. Together, our contributions help demystify the latent spaces of language models and offer new insights into how to harness model internals to build more trustworthy language technology.

Nishant Subramani

Carnegie Mellon University, Language Technologies Institute

Language Model Steering LLM Interpretability Language Model Calibration LLM Reliability Activation Steering

Ask PDF

Mar 30, 2026

Variational Neurons in Transformers for Language Modeling

Transformers for language modeling usually rely on deterministic internal computation, with uncertainty expressed mainly at the output layer. We introduce variational neurons into Transformer feed-forward computation so that uncertainty becomes part of the internal computation itself. Concretely, we replace deterministic feed-forward units with local variational units based on EVE while preserving the overall Transformer backbone. We evaluate this design in compact next-token language-modeling settings. We compare deterministic and variational variants with both predictive and probabilistic criteria. Alongside negative log-likelihood, perplexity and accuracy, we analyze calibration, conditional variance, mutual information and latent-usage statistics. The resulting picture is clear. Variational neurons integrate stably into Transformers, preserve strong predictive performance and produce informative uncertainty signals. The experiments also show that task quality, useful depth and internal stability are distinct properties. These results establish variational Transformers as a practical form of uncertainty-aware language modeling. They show that Transformers can predict with an explicit internal structure of uncertainty, which supports stronger probabilistic evaluation and a more informative analysis of model behavior.

Yves Ruffenach

Conservatoire National des Arts et Métiers Strasbourg, France

LLM Uncertainty Estimation Transformer Transformer FFNs Autoregressive Language Modeling

Ask PDF

May 19, 2026

Reading Calibrated Uncertainty from Language Model Trajectories

The maximum softmax probability (MSP) represents a default approach when evaluating uncertainty quantification for language model generation with structured output. Although cheap, it is often miscalibrated. Methods that probe the model's internal activations feed raw hidden states into opaque classifiers, reading activations as static snapshots and leaving implicit the layer-wise trajectory by which a representation is formed. Yet, similar endpoints can arise from very different paths, and how evidence accumulates, reinforces, or reverses across depth might reveal uncertainty that final probabilities obscure. We extract eleven scale-invariant geometric features, tracing the cumulative path of per-layer MLP updates, and feed them to a sparse linear probe. The probe outperforms MSP under selective abstention, with gains scaling with baseline miscalibration up to 21 AURC points. Because every feature has a closed-form geometric meaning, the probe's coefficients trace how and where along depth errors take shape -- which layers commit prematurely, which contradict the running state, where trajectories drift away from their endpoint.

Aliai Eusebi, Alexander Herzog, Xiaoyu Liang +3

University College London, London, United Kingdom · Core46, London, United Kingdom

Confidence Estimation in Language Models LLM Reliability LLM Uncertainty Estimation Representation Geometry in Language Models Uncertainty Calibration Language Model Probing

Ask PDF