A model compressed to fit on a phone or edge device has two problems at once. Compression has already damaged it, and in deployment it meets data unlike its training set. Recent work shows that a single yes/no bit per example (a user or checker saying whether the prediction was right) is enough to adapt a full-precision image classifier to such shifts. That bit beats both unsupervised self-training and asking for the full correct label. Separate work shows that a compressed model's damage can be partly repaired by learning a handful of scalars along a few fixed directions, the directions in which its internal activations deviate from the uncompressed original. We propose joining the two. A compressed model should spend its scarce yes/no feedback adjusting only those repair scalars, and it should decide which examples to ask about using the gap between its compressed and repaired views rather than expensive repeated stochastic forward passes.
The bet is that, for a compressed model, the directions along which compression hurt it are also where shift-induced errors concentrate. If that holds, restricting test-time learning to a tiny compression-aware subspace should make learning from yes/no feedback both cheaper and more stable than adapting generic parameters. The bet can fail, and the program is built so that a failure is visible.
Core Claim
BiTTA adapts a model under domain shift by querying a yes/no oracle on uncertain samples. It rewards correct predictions with +1 and incorrect ones with −1 through REINFORCE, a policy-gradient rule that needs only a scalar reward per sample. Confident unqueried samples get self-reward to prevent forgetting. BiTTA reports a 13.3 percentage-point average gain over TENT, SAR and CoTTA on CIFAR10-C, CIFAR100-C and Tiny-ImageNet-C, and up to 32% higher accuracy than full-label active adaptation at equal labeling cost.
Two limits matter here. BiTTA picks samples by epistemic uncertainty from Monte Carlo Dropout, which means several stochastic forward passes per input, and the authors say this is too costly for edge hardware. It also assumes a full-precision backbone. Compressed models are exactly where calibration breaks: Zhang et al. find a cliff at 3 bits, with WebQuestionsSP accuracy at 12.52% and rising hallucination. This creates a tension. The model most in need of feedback is the one whose uncertainty estimates are least trustworthy for deciding when to ask for it.
Our claim is that both limits can be addressed with a structure borrowed from compression repair. RestoreLCC contrasts the activations of dense and pruned models, takes a singular value decomposition (SVD) of the deficit at the most-affected attention heads, freezes the resulting directions, and learns one scalar per direction. These scalars fold into existing bias terms, adding zero latency and under 0.05% parameters, and give +2.01% mean accuracy on 50%-pruned LLaMA-7B. We keep those directions and learn their scalars online from yes/no rewards. We also replace Monte Carlo Dropout with a two-view disagreement signal: the compressed model's prediction with and without the current repair. That costs one extra pass instead of many.
To make this concrete, let $z_\ell$ be the activation at layer $\ell$, $u_{\ell k}$ the $k$-th frozen deficit direction from offline SVD, and $\beta_{\ell k}$ its learned scale. The repaired model computes $z_\ell + \sum_k \beta_{\ell k} u_{\ell k}$. The online update is BiTTA's REINFORCE step restricted to the vector $\beta$:
$$\beta \leftarrow \beta + \eta \, r \, \nabla_\beta \log p_\beta(\hat y \mid x),$$
where $r \in \{+1,-1\}$ is the yes/no bit, $\hat y$ the predicted label, and $\eta$ a step size. The gradient flows through frozen dequantized weights, so the quantized weights themselves are never updated.
Research Program
The first stage is offline and done once before deployment. We take a source-trained network, compress it, and extract per-layer deficit directions on clean source data following RestoreLCC. We cover two compression types because LLMCBench reports that they fail differently: pruning preserves reasoning but loses facts, while quantization does the opposite. The first is post-training weight quantization at 4, 3 and 2 bits. The second is 50% unstructured pruning, with the caveat from LLMCBench that unstructured sparsity rarely yields real speedups, so pruning results are reported for accuracy only.
The second stage is online. A stream of shifted test data arrives. For each input the device computes the repaired prediction and the unrepaired prediction, and it queries the oracle when they disagree or when their gap is large. The query threshold follows the logic of opportunistic scheduling in wireless networks, where Niu et al. and Hassel et al. broadcast a threshold and let only users above it reply with one bit. In our case the threshold is adjusted online so that the realized query rate tracks a fixed labeling budget. The repair scalars are then updated from the yes/no bit, and BiTTA's agreement-based self-reward is applied to confident, unqueried samples.
The third stage is analysis. The question is whether shift-induced errors actually lie in the compression-deficit subspace. For queried errors, we measure how much of the gradient with respect to all bias-like parameters projects onto the deficit directions, compared with random directions of equal count. This tests the central hypothesis directly rather than inferring it from accuracy alone.
A secondary, smaller extension uses 50%-pruned and 3-bit LLaMA-7B on BoolQ and PIQA, where the yes/no bit is answer correctness. This checks whether the mechanism transfers beyond vision.
Baselines
The naive reference point is the compressed model with no adaptation, alongside the same model with offline RestoreLCC repair fitted on clean source data and then frozen. The second variant tells us whether online feedback adds anything beyond offline repair. Both are run on identical compressed checkpoints. No reported result was found on Lacuna for either on CIFAR-C under quantization.
The strongest competitor is BiTTA itself, run unchanged on the same compressed backbone. It updates the same normalization and affine parameters as in its paper and selects queries with Monte Carlo Dropout. This is the method that already shows yes/no feedback works, so any gain must come from our compression-specific choices. The only reported reference figure is BiTTA's 13.3-point average gain over TTA baselines on full-precision CIFAR10-C, CIFAR100-C and Tiny-ImageNet-C. No reported result was found on Lacuna for BiTTA on quantized or pruned backbones.
Fairness rules apply to every comparison. Methods share the same stream order, the same number of oracle queries per sample, and the same hyperparameter search budget per method. We report the forward passes each method spends per input. Beating BiTTA means equal or higher final accuracy and expected calibration error (ECE) at the same query budget, with at most one extra forward pass per input against BiTTA's several. Alternatively, it means a clear accuracy gain at 3 and 2 bits, where we expect BiTTA's uncertainty estimates to degrade.
Unsupervised test-time adaptation (TENT and CoTTA on the compressed model) sets the zero-label floor. Full-label active adaptation at equal labeling cost sets the richer-supervision reference.
Two ablations isolate the key ingredients. The first keeps our query rule but swaps the deficit directions for random directions with the same parameter count. It also tests BiTTA's normalization parameters and a LoRA adapter of matched size, which isolates the update subspace. The second keeps the deficit-direction update but swaps the query rule for Monte Carlo Dropout or for Distance-to-Uniform confidence from Varshney et al., a zero-extra-pass score. This isolates the query trigger.
Why not just…?
Why not just run BiTTA on the compressed model as-is? This is the main threat, and it is the head-to-head baseline. BiTTA's own paper says Monte Carlo Dropout is too expensive on edge hardware, so it fails the deployment setting on cost even if it holds up on accuracy. Whether its accuracy holds at 3 and 2 bits is unknown: no reported result was found on Lacuna. If BiTTA matches us at 2 bits with equal passes, our subspace claim is refuted.
Why not just use fewer bits of compression, say 4-bit instead of 3-bit? LLMCBench reports that quantization keeps over 95% of performance at moderate settings, so a milder compression level may make adaptation unnecessary. We therefore compare at matched memory. A 3-bit model with online repair is set against an unadapted 4-bit model, and against 4-bit BiTTA, on the same shifted streams. If the 4-bit model with no feedback wins, the case for adapting at 3 bits fails in this setting.
Why not just use unsupervised self-adaptation and skip the oracle? BiTTA shows entropy minimization accumulates its own mistakes under severe shift, losing 13.3 points on average. We expect this to be worse on miscalibrated compressed models. TENT and CoTTA on the compressed backbone test this directly.
Why not just fine-tune a LoRA adapter, or a soft prompt as in Compress, Then Prompt, from the yes/no rewards? Both are reasonable carriers of the update. RestoreLCC reports that LoRA struggles at high sparsity and adds latency when left unmerged. Soft prompts are relevant only to the language extension. The matched-size subspace ablation answers this empirically. If LoRA learns faster from the bits than the deficit directions do, we adopt LoRA, and the contribution reduces to the query rule.
Proposed Experiments
The main vision study uses ResNet and ViT backbones on CIFAR10-C, CIFAR100-C and Tiny-ImageNet-C, with ImageNet-C and PACS as harder shifts. These are BiTTA's benchmarks, so its protocol and labeling-cost accounting carry over. Backbones are compressed to 4, 3 and 2 bits and to 50% sparsity. We report online accuracy, ECE, queries per sample, forward passes per input, and memory.
Three checks decide the outcome. The first is the subspace projection: the deficit directions should capture clearly more of the error gradient than random directions, and if they do not, the core hypothesis is false regardless of accuracy. The second is the 2-bit head-to-head against BiTTA at equal queries. The third is whether disagreement-based querying still selects mostly wrong predictions at 2 bits, which we measure as the precision of queries at catching errors, while Monte Carlo Dropout and Distance-to-Uniform lose that ability. The language extension on 50%-pruned LLaMA-7B (BoolQ, PIQA) is scored against RestoreLCC's reported +2.01% offline gain. It tests whether online yes/no feedback adds accuracy beyond what offline repair already provides.
References
- Test-Time Adaptation with Binary Feedback
- Toward Democratized Generative AI in Next-Generation Mobile Edge Networks
- Restoring Pruned Large Language Models via Lost Component Compensation
- LLMCBench: Benchmarking Large Language Model Compression for Efficient Deployment
- Ergodic and Outage Performance of Fading Broadcast Channels With 1-Bit Feedback
- Rate-Optimal Multiuser Scheduling with Reduced Feedback Load and Analysis of Delay Effects
- Model Cascading: Towards Jointly Improving Efficiency and Accuracy of NLP Systems
- Compress, Then Prompt: Improving Accuracy-Efficiency Trade-off of LLM Inference with Transferable Prompt