When labels are expensive, active learning lets the model choose which examples a person should annotate next. The usual rule picks the examples the model is least sure about. That rule has a known failure: the least certain examples are often outliers, such as garbled or off-topic inputs, and labeling them teaches the model almost nothing about the rest of the data. The standard fix multiplies uncertainty by a "density" term, so that the chosen examples are both confusing and typical of many other examples. Today that density is computed from the distances between frozen pre-trained sentence embeddings. Meanwhile, the cheap way to fine-tune large models under label scarcity is low-rank adaptation (LoRA). LoRA freezes the pre-trained weights $W_0$ and learns a small correction $BA$ of rank $r$, so that a layer computes $h = W_0x + BAx$. Embedding distances say nothing about which examples would move those few trainable parameters in the same direction.
This note proposes measuring density in the space of LoRA gradients. Two unlabeled examples count as neighbors if labeling either one would push the adapter the same way. The candidate to label is then one that is uncertain and whose adapter-gradient neighborhood is crowded, so a single label resolves the error on many similar inputs at once. LoRA makes this practical in two ways. The gradient with respect to $B$ is an outer product of two small vectors, so similarities between gradients factor into cheap pieces. And because $B$ starts at zero, its gradient is the one gradient that is informative at the very first round, when adapters are untrained and other uncertainty signals are noise.
Core Claim
The closest existing work repairs uncertainty or repairs density, but never measures density in the geometry the adapter actually trains. STAR shows that plain active learning with LoRA on LLaMA2-7B can lose to random selection. Maximum entropy gets a relative improvement over passive learning of −0.22 on OpenBookQA. STAR traces this to untrained adapters and to overconfidence, with about 70% confidence on wrong answers. It fixes the problem by blending the frozen model's uncertainty with the adapter's, putting weight decay on $B$, and applying Monte-Carlo dropout on $A$. Its acquisition score is still pure uncertainty, with no protection against outliers. Jukic and Snajder show that parameter-efficient tuning beats full fine-tuning by 5–10 F1 points at 100 or fewer labels on SUBJ, TREC, SST and AGN. They only test off-the-shelf entropy and core-set selection. Density-aware methods supply the other half of the picture. These are Settles and Craven's information density, contrastive active learning (CAL), DEUCE and vote-k. All of them compute neighborhoods in fixed embedding spaces, and none runs inside a multi-round LoRA loop.
The claim is specific. Take STAR's calibrated uncertainty and multiply it by a local density computed from adapter-gradient directions. This should beat the same uncertainty multiplied by embedding-space density. The advantage should be largest in the earliest rounds and on pools that contain outliers or tightly packed but task-irrelevant regions. The design anticipates one way the claim could fail. If frozen embedding density does just as well, the obvious combination is enough, and the paper reports that.
Research Program
For an unlabeled example $x$, write $a_\ell(x) = A_\ell x_\ell$ for the $r$-dimensional projection computed by LoRA module $\ell$. Write $\delta_\ell(x)$ for the gradient of the loss with respect to that module's output, using the model's own predicted label. The gradient with respect to $B_\ell$ is then $\delta_\ell(x)\,a_\ell(x)^\top$. Pooling over tokens gives the similarity between two examples in closed form:
$$k(x,u) = \sum_{\ell \in L} \langle \delta_\ell(x), \delta_\ell(u)\rangle \, \langle a_\ell(x), a_\ell(u)\rangle,$$
Here $L$ is the set of the last few LoRA modules. The kernel never builds the $d \times r$ gradient matrices. The acquisition score is
$$\text{Score}(x) = U(x)\cdot\Big(\tfrac{1}{k}\textstyle\sum_{u \in N_k(x)} \bar k(x,u)\Big)^{\beta},$$
where $U(x)$ is STAR's dynamic Monte-Carlo-dropout uncertainty, $\bar k$ is $k$ normalized to cosine form, $N_k(x)$ is the $k$ nearest pool neighbors under $\bar k$, and $\beta$ trades off the two terms as in Settles and Craven.
Two design choices matter. First, the density is local, computed over $k$ neighbors, rather than Settles' average over the whole pool. With cosine similarity, a whole-pool average reduces to a dot product with the mean vector, which measures global centrality rather than cluster membership. Neighbors are found by approximate nearest-neighbor search over a tensor sketch of the factored vectors, so the cost is near-linear in the pool size rather than the $O(N^2)$ the file identifies as the bottleneck. Second, the gradient direction is normalized, so it carries no confidence information. Confidence stays in $U(x)$. This prevents STAR's overconfidence problem from leaking into the density term.
Why expect gradient geometry to help at round zero? With $B=0$, the gradient with respect to $A$ vanishes, but the gradient with respect to $B$ equals $\delta \otimes Ax$. That is the frozen model's error signal crossed with a random projection of the layer input, and a random projection approximately preserves distances. So the density is well-defined and task-aware before any adapter training. This speaks directly to the "under-fitted early adapters" problem that STAR handles only by down-weighting adapter uncertainty. Beyond the method, the program includes a diagnostic. It measures how well each density estimate predicts the actual held-out gain from labeling an example, which tests whether "density" in either space tracks what it is meant to track.
Baselines
The naive reference point is uniform random selection with the same LoRA setup, retrained each round. The file shows this is not a strawman: entropy loses to it on OpenBookQA. Beating random by a margin larger than across-seed variation is the minimum bar.
The strongest existing method is STAR, which reports a relative improvement over passive learning (RIPL) of 7.47 on OpenBookQA with LLaMA2-7B. It also reaches peak BoolQ accuracy at 1,000 labels, though the file gives no numeric BoolQ accuracy. STAR will be rerun with its own regularization, uncertainty schedule, rank, and per-round budget. Our method changes only the acquisition score. Beating STAR means a higher RIPL on OpenBookQA and BoolQ averaged over at least five seeds, with confidence intervals that do not overlap 7.47's reproduced counterpart.
The key ablation is STAR uncertainty multiplied by $k$-nearest-neighbor information density computed in frozen sentence-embedding space, with the same $k$, $\beta$, and nearest-neighbor machinery. This is the obvious combination. It isolates the single ingredient the proposal adds, namely the space in which density is measured. A second ablation removes uncertainty entirely and selects by gradient-space density alone, to test whether the product matters.
On the encoder benchmarks, the comparisons are CAL and DEUCE seed selection, both run with LoRA instead of full fine-tuning, plus the entropy and core-set strategies from Jukic and Snajder. These run inside their released adapter-al codebase on BERT-base, so the data splits, rounds, and adapter configurations match exactly. For the encoder setting the file contains no numeric F1 or RIPL. The relevant facts are the 5–10 point gap over full fine-tuning and DEUCE's reported 6.2% gain on TREC and 2.5% on balanced datasets over standard heuristics. Every method gets the same hyperparameter-tuning budget on a held-out labeled development set of fixed size. The extra backward passes our method needs are reported as wall-clock cost per round.
Why not just…?
Why not just multiply STAR's uncertainty by embedding-space density? This is the real threat, and it is the central ablation rather than an afterthought. The evidence it might fall short is indirect. The file notes that pre-trained embeddings can be anisotropic and may miss domain nuances, so dense embedding regions need not be dense in task-relevant directions. Jukic and Snajder also show that adapters reshape late-layer representations while early ones stay fixed. If the gradient-space version does not beat this ablation on the round-by-round learning curve and on the outlier-injection test, the paper's conclusion becomes that the simple combination suffices. That would still be the first test of density-aware LoRA active learning.
Why not just use random selection with LoRA, since the low-rank bottleneck already regularizes? Random selection beats naive uncertainty sampling, as STAR's −0.22 shows. But calibrated selection beats random, as STAR's 7.47 shows, so selection quality clearly matters. Random sampling is included at every budget. If our method's margin over random shrinks to nothing after a few hundred labels, we will report the budget range where selection pays off.
Why not just pick a representative set once with vote-k or DEUCE and skip uncertainty? DEUCE already shows that pure geometric coverage misses minority classes on TREC. vote-k was built for in-context prompting and never tested with LoRA. Both are run as one-shot seed selectors followed by LoRA training and as full selectors. The class-imbalance metric on TREC will show whether a static representative set eventually stalls.
Proposed Experiments
Two settings are used. The encoder setting is BERT-base with LoRA on SUBJ, TREC, SST and AGN, using the protocol and code of Jukic and Snajder, with budgets from 50 to 1,000 labels. The decoder setting is LLaMA2-7B with STAR's protocol on BoolQ and OpenBookQA, plus GSM8K, where the token-level loss defines $\delta$. Metrics are RIPL, area under the learning curve, macro-F1 or accuracy per round, and the label-imbalance metric on TREC.
Three checks are decisive. The first is an outlier-injection test. We add 5% corrupted pool examples, made of token-shuffled text and off-domain sentences, and count how many each method queries. Pure STAR should take many of them. Both density variants should take fewer, and the gradient variant should take the fewest if its premise is right. The second is a calibration-of-density test. On a subsample of a few hundred candidates per round, we actually add each candidate to the training set, retrain the adapter for the same number of steps, and record the held-out gain. We then compare the rank correlation of that gain with gradient-space density against its correlation with embedding-space density. The claim stands or falls on gradient density correlating more strongly, especially in rounds one and two. The third is a round-zero test. With adapters untrained, we compare selections made by gradient density alone against DEUCE and vote-k seed sets on TREC and SST. This tests whether the $B$-gradient really is informative before training, as the analysis predicts.
References
- STAR: Constraint LoRA with Dynamic Active Learning for Data-Efficient Fine-Tuning of Large Language Models
- Parameter-Efficient Language Model Tuning with Active Learning in Low-Resource Settings
- An Analysis of Active Learning Strategies for Sequence Labeling Tasks
- Active Learning by Acquiring Contrastive Examples
- Deuce: Dual-diversity Enhancement and Uncertainty-awareness for Cold-start Active Learning
- Selective Annotation Makes Language Models Better Few-Shot Learners