Single-Pass Uncertainty for Proactive Agents on Untrusted Edge Silicon

A wearable assistant that speaks up only when it should is still mostly a research dream. When large models try to help proactively, they interrupt constantly: on the ProactiveBench benchmark of streaming user-activity events, GPT-4o-mini fires a false alarm on 64.73% of the moments where it should have stayed silent, and even the best fine-tuned open model reaches only 66.47% F1 on deciding when to intervene (Proactive Agent: Shifting LLM Agents from Reactive Responses to Active Assistance). The standard fix — estimate how uncertain the model is before acting, using Monte Carlo dropout, i.e. running the same network $N$ times with random units disabled and looking at the spread of answers (Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning) — multiplies inference cost by $N$ on every single frame, which a passively cooled wearable cannot afford. There is a second, quieter problem stacked on top: the assistant's perception encoder runs on a chip fabricated by an untrusted offshore foundry, and hardware-security research has shown such silicon can carry dormant modifications that pass all standard tests and activate only on rare triggers (TrojanSAINT: Gate-Level Netlist Sampling-Based Inductive Learning for Hardware Trojan Detection).

This proposal claims that one number, computed in a single forward pass, can do three jobs at once: decide when the agent should stay silent, decide how many bits it should send upstream, and help detect when its own chip is lying. The number is the compression rate of an Information Bottleneck encoder — the KL divergence between the encoder's belief about its compressed representation and a fixed prior — which Uncertainty in the Variational Information Bottleneck showed is a usable uncertainty signal without any repeated sampling. We extend that observation from isolated image classification into a full proactive-agent pipeline, and we confront a tension nobody has examined: the same compression that makes the uplink cheap also throws away the subtle statistical anomalies that runtime Trojan detectors rely on.

Core Claim

Let the edge device encode its sensor context $x$ into a latent code $z$ whose distribution $q(z \mid x)$ is trained with a Variational Information Bottleneck objective — keep only what the downstream task needs, discard the rest. The training procedure already computes

$$R(x) \;=\; \mathrm{KL}\big(q(z \mid x)\,\|\,r(z)\big),$$

where $r(z)$ is a fixed prior, and this quantity is available at test time in one forward pass. We claim $R(x)$ can simultaneously serve as (i) a silence gate — the agent intervenes only when predicted task utility is high and $R(x)$ sits in a calibrated band, suppressing the false alarms that plague eager agents; (ii) a rate controller — the number of transmitted symbols grows with uncertainty,

$$k(x) = k_0 + \lceil \gamma \, R(x) \rceil,$$

extending the variable-length feature transmission of Learning Task-Oriented Communication for Edge Inference: An Information Bottleneck Approach, which varied length with channel quality but not with epistemic uncertainty; and (iii) the driver of a small, fixed-budget attestation stream described below.

The third job addresses an unresolved conflict in the literature. Task-oriented encoders are explicitly optimized to discard nuisances — background texture, lighting, high-frequency detail (Learning Task-Oriented Communication for Edge Inference: An Information Bottleneck Approach) — while runtime defenses against Trojaned neural hardware, such as CLEANN's frequency-domain and sparse-recovery analyzers, detect dormant triggers precisely through subtle anomalies in activations (CLEANN: Accelerated Trojan Shield for Embedded Neural Networks). An aggressive compressor may therefore strip the evidence a server-side detector needs. Our answer is not to weaken compression but to split the uplink: the task path stays maximally compressed, while a small protected side-channel carries quantized random projections of intermediate encoder activations, chosen so that rare trigger-induced perturbations remain statistically visible. This is deliberately statistical attestation, not cryptographic proof — Verifiable ASICs: Bridging the Trust-Performance Gap in Semiconductor Supply Chains demonstrates that true verifiable computation needs hardware–protocol co-design and still only handles wide, shallow arithmetic circuits, with general-purpose schemes costing $10^5\times$ to $10^7\times$ overhead, far beyond deep network inference. We trade Zebra's ironclad guarantees for something that fits a wearable's budget, and we measure exactly what that trade buys.

Research Program

We will build a three-part system. First, an edge encoder — a VIB-trained feature extractor with two heads: a small utility predictor estimating whether proactive help would be valued, and the rate computation above. Second, the attestation head: a fixed random projection matrix applied to intermediate activations, quantized to a few bits per symbol and appended to the uplink at constant budget regardless of $k(x)$. Third, a server side with a task head, a CLEANN-style anomaly analyzer running on the received stream, and a conformal anomaly test on the attestation symbols in the spirit of Conformal Semantic Communication: Distribution-Free Task-Level Coverage Guarantees for Goal-Oriented Transmission Under Channel Shift, which used importance-weighted conformal scores to keep task-level coverage within ±0.8% of a 90% target under channel shift at roughly 0.7 ms per sample.

Alongside the system we will characterize the compression–detectability tradeoff empirically: sweep the bottleneck strength, inject trigger patterns of varying frequency content, and map which trigger signatures survive (a) the IB pruning, (b) quantized random projection, and (c) the wireless channel. This is the analytical heart of the work — no one has measured whether nuisance-suppression and tamper-evidence are fundamentally at odds, or only accidentally so. The security model follows the latent-manipulation threat formalized in Rethinking Security in Semantic Communication: Latent Manipulation as a New Threat, which showed an adversary can steer semantic latents while keeping them statistically natural — but moved from a network man-in-the-middle to a silicon-level source: the untrusted foundry that fabricated the encoder chip, modeled behaviorally via the BadNets/TrojanNN trigger suites used by CLEANN and, for structural plausibility, the dormant low-controllability trigger patterns documented across the 17 TrustHub benchmarks in TrojanSAINT.

Baselines

The naive reference point is an ungated proactive agent: on ProactiveBench, that means the off-the-shelf models evaluated in the Proactive Agent paper, where GPT-4o-mini's 64.73% false-alarm rate defines the pain threshold, and where the strongest published system, fine-tuned Qwen2-7B-Proactive, reaches 66.47% F1 — the number to beat or match. The strongest uncertainty-based competitor is Monte Carlo dropout gating: the same backbone, same fine-tuning data, ten stochastic passes per event, variance threshold tuned with the same budget we give our own gate. This is the fairest test of the single-pass claim because it holds everything equal except where the uncertainty estimate comes from. A second strong competitor is a softmax-confidence gate on the same fine-tuned model — cheap, and the default a practitioner would actually deploy. On the tamper side, the strongest existing runtime defense is CLEANN relocated to the server, running its frequency-domain and sparse-recovery analysis on whatever the uplink delivers; on its home turf it reports reducing attack success rate to 0% across its attack suite with only a 0.1% clean-accuracy drop on MNIST, so the question is not whether CLEANN works but whether compression starves it. The design-time alternative is a TrojanSAINT-style netlist audit, which reports 78% true-positive and 85% true-negative rates on unseen Trojans without golden references — relevant as an upper bound on what pre-silicon screening alone achieves, since it cannot see runtime behavior at all. Finally, two ablations isolate our ingredients: rate-only gating with no attestation channel, and attestation-only with fixed-rate transmission. Beating the baselines means, on ProactiveBench, F1 at or above 66.47% with a false-alarm rate clearly below both the ungated and softmax-gated models at single-pass cost; on the tamper testbed, attack success rate near CLEANN's 0% while transmitted bits stay within roughly 15% of the unprotected compressed uplink.

Why not just…?

Why not just threshold the softmax confidence? This is the most serious alternative, and the evidence against it is only partial. Dropout as a Bayesian Approximation showed deterministic networks are overconfident — a model can output 99% softmax probability on inputs it has never seen — and the VIB paper found deterministic baselines push probabilities toward 0 or 1, with temperature scaling matching the rate metric's ~91% AUROC on in-distribution error detection; the rate metric's edge appears only on out-of-distribution inputs, where it reached 89.7% AUROC on flipped images while entropy-based detection lagged. Since ambiguous proactive contexts are often distribution-shifting, we expect the gap to matter — but we will run the softmax gate as a first-class baseline, and if it ties our gate on ProactiveBench we will report that honestly and rest the proposal's value on the rate control and attestation jobs.

Why not just run MC dropout with ten passes? It is the incumbent method and plausibly gates well. The cost argument is real — AdaNSP: Uncertainty-driven Adaptive Decoding in Neural Semantic Parsing shows even adaptive systems still burn multiple passes just to decide whether to think harder, and the direction survey notes 10–50 passes per frame blows sub-10ms latency budgets and thermally throttles wearables — but the file contains no measured per-pass latency for MC dropout on edge hardware, so we will not assert a number; Experiment C measures it directly on our platform at equal tuning budget. If ten-pass gating buys several F1 points over the single-pass gate, the honest conclusion is a hybrid: single-pass gating with occasional multi-pass escalation, and we would report that frontier rather than defend the pure version.

Why not just use standard codecs, AES encryption, SHA integrity checks, and a pre-silicon netlist audit? Each piece fails at a specific joint. Bit-exact codecs hit cliff effects under channel degradation — Vector Quantized Semantic Communication System shows BPG's quality collapses at a threshold about 1 dB higher than its learned alternative in both AWGN and Rician channels — and transmitting reconstructed pixels wastes bandwidth that language-style semantic tokens cut by 96% in LangCoop: Collaborative Driving with Language. Cryptographic hashes verify the bits, not the meaning: the latent-manipulation paper demonstrates attacks whose outputs look statistically natural and pass bit-level scrutiny. And the netlist audit is blind by construction — dormant triggers on low-controllability signals evade functional test (TrojanSAINT) — leaving exactly the runtime window our attestation stream covers. The bundle is not silly, though; Experiment B includes it, because if a hash plus audit catches our injected triggers, the attestation channel is not earning its bits.

Proposed Experiments

Experiment A tests the silence gate on ProactiveBench — 6,790 training events across coding, writing, and daily life, 233 real-world test events, with 1,640 human-labeled reward entries whose associated reward model reaches 91.8% agreement with human accept/reject judgments (Proactive Agent). We fine-tune a Qwen2-7B-class backbone with a VIB bottleneck on the context encoder and gate interventions with the utility head plus $R(x)$, then compare against the ungated fine-tuned model (published reference: 66.47% F1; GPT-4o at 64.60% and Claude-3.5-Sonnet at 62.00%; LLaMA-3.1-8B rising from 55.06% to 66.25% after training on this data), a softmax gate, and MC-dropout gating at $N \in \{5, 10\}$, all given the same fine-tuning data and threshold-tuning budget. Metrics are F1, false-alarm rate, and measured per-event latency and energy. Success is F1 at or above 66.47% with false alarms well under the 64.73% naive rate at single-pass cost; a tie between softmax and rate gating would itself be an informative negative for the uncertainty-gating literature.

Experiment B tests the tamper story. We build a device-edge pipeline over CIFAR-10 and GTSRB (the latter from CLEANN's attack suite, alongside MNIST), sending VQ-style compressed features over simulated AWGN and Rayleigh channels across the −5 dB to 20 dB sweep used by the conformal semantic-communication paper. Into the encoder we inject BadNets and TrojanNN triggers, and we measure attack success rate, clean accuracy, bits per sample, and detector AUROC under four conditions: no defense; CLEANN at the server reading only the compressed task latents (our gap hypothesis predicts its 0%-ASR performance degrades as the bottleneck tightens); the full attested uplink; and the codec-plus-hash-plus-audit bundle. The decisive plot is detection AUROC versus bits per sample as we sweep the bottleneck strength: if the attestation channel shifts this frontier outward, the compression–detectability tension is real but manageable; if the frontier barely moves, the tension is fundamental and that finding is itself the contribution. CLEANN's published 0% ASR with 0.1% clean drop on MNIST is the reference point the attested system must approach while the pure task path stays within about 15% of the unprotected uplink's bit budget.

Experiment C is the overhead accounting: single-pass rate gating versus MC dropout at $N$ up to 50 on identical hardware, reporting wall-clock and energy per event, benchmarked against the roughly 0.7 ms per sample the conformal approach reports; the file contains no comparable measured figure for multi-pass dropout on edge hardware, so we will produce one. The decisive checks that could kill the idea are explicit: if softmax gating matches rate gating on ProactiveBench, if CLEANN survives aggressive compression unaided, or if the attestation bits needed for reliable detection exceed the savings from compression, the proposal's core claim fails in its current form — and each experiment is designed so that failure would be visible and reportable, not buried.