Chat-CBM: Conversational Concept Bottlenecks

• Chat-CBM is a concept bottleneck model that replaces traditional score-based classifiers with semantic reasoning via a frozen LLM.
• It decodes numerical concept activations into textual descriptions to facilitate interactive interventions and refined label predictions.
• Empirical results demonstrate improved accuracy and robustness, particularly in fine-grained and unsupervised settings.

Searching arXiv for the Chat-CBM paper and closely related concept bottleneck work. Chat-CBM is a concept bottleneck model in which the conventional score-based label predictor is replaced by a language-based classifier implemented with a frozen LLM. In standard CBMs, an input $x \in \mathcal{X}$ is mapped to a concept activation vector $c \in [0,1]^{N_c}$, and then to a class label $y \in \{1,\dots,M\}$ through a typically linear classifier. Chat-CBM preserves the bottleneck structure but shifts prediction into the semantic space of concepts: activated concepts are decoded into textual concept semantics, candidate labels are proposed by the original CBM head, and a frozen LLM reasons over those semantics, optionally augmented by in-context demonstrations and class-prior descriptions, to select a label (He et al., 22 Sep 2025).

1. Origins in concept bottlenecks and interactive prediction

Concept bottleneck models are interpretable neural networks that first predict labels for human-interpretable concepts relevant to the prediction task and then predict the final label from those concept predictions. In the standard formulation, the concept predictor $g$ produces $\hat c = g(x)$, with $g_j(x) \in [0,1]$ for concept $j$, and the label predictor $h$ computes

$$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$

This architecture makes the internal decision interface legible and editable, but the usual fixed linear classifier restricts interventions to numerical manipulation of concept scores (He et al., 22 Sep 2025).

A precursor line extended CBMs to interactive prediction settings by allowing the model to query a human collaborator for the label to some concepts at test time. In that setting, the model maintains a set $S\subset\{1,\dots,m\}$ of acquired ground-truth concepts and chooses which remaining concept to request next so as to maximally improve the final prediction. The policy combines concept prediction uncertainty and concept influence on the current top-scoring label, and the resulting interactive CBM was reported to achieve accuracy gains of $c \in [0,1]^{N_c}$0–$c \in [0,1]^{N_c}$1 with only $c \in [0,1]^{N_c}$2 interactions over competitive baselines on Caltech-UCSD Birds, CheXpert, and OAI (Chauhan et al., 2022).

This progression clarifies Chat-CBM’s research position. Interactive CBMs made the bottleneck queryable; Chat-CBM generalizes the intervention interface by replacing a fixed score-based classifier with a conversational, semantics-aware classifier. This suggests that the central innovation is not the bottleneck itself, but the relocation of decision making from a linear map over scores to an LLM that reasons over concept names, descriptions, and user-supplied language.

2. Semantic bottleneck architecture

Chat-CBM introduces what the paper describes as a language-based classifier on a semantic bottleneck. Rather than sending the numerical vector $c \in [0,1]^{N_c}$3 directly into a fixed linear layer, the model decodes $c \in [0,1]^{N_c}$4 into a semantic list

$c \in [0,1]^{N_c}$5

consisting of the textual names or short descriptions of all concepts whose score exceeds $c \in [0,1]^{N_c}$6, or the top-$c \in [0,1]^{N_c}$7 concepts in unsupervised settings. The original label head $c \in [0,1]^{N_c}$8 is still used, but only to generate a candidate set

$c \in [0,1]^{N_c}$9

The final prompt is a knowledge-enhanced semantic bottleneck prompt comprising four components: a fixed task instruction $y \in \{1,\dots,M\}$0, an in-context demonstration set $y \in \{1,\dots,M\}$1, an optional class-prior $y \in \{1,\dots,M\}$2 describing common attributes for each candidate class, and the test-time semantic list $y \in \{1,\dots,M\}$3. The LLM is then asked to choose one of the candidates, with the paper formalizing the decision rule as

$y \in \{1,\dots,M\}$4

In this construction, the original CBM still supplies the bottleneck and a candidate shortlist, but semantic reasoning is delegated to the frozen LLM (He et al., 22 Sep 2025).

The architecture is motivated by a specific limitation of traditional CBMs: users can intervene in the concept space, but a fixed linear classifier over concept scores prevents the incorporation of new concepts or domain knowledge at test time. The 2025 paper further states that this limitation is particularly severe in unsupervised CBMs, where concept activations are often noisy and densely activated, making user interventions ineffective. Chat-CBM addresses this by grounding prediction in the semantic space of concepts rather than only in their numerical activations (He et al., 22 Sep 2025).

3. Intervention interface

Because the label predictor is conversational, Chat-CBM supports two classes of intervention. The first is numerical concept correction, which mirrors standard CBMs. A user edits the concept activation vector, $y \in \{1,\dots,M\}$5, after which the model recomputes

$y \in \{1,\dots,M\}$6

and reinvokes the LLM to obtain an updated prediction

$y \in \{1,\dots,M\}$7

The second class is conversational intervention. The model maintains a chat history $y \in \{1,\dots,M\}$8 and accepts free-form textual feedback $y \in \{1,\dots,M\}$9. The paper enumerates three types: concept correction, concept addition or removal, and high-level reasoning guidance. The LLM input becomes

$g$0

and the updated prediction is

$g$1

Illustrative interventions given in the paper include “Actually the bird’s bill is hooked, not straight,” “Add concept: ‘yellow belly’,” “Ignore size-related concepts,” and “Focus on feather pattern when uncertain” (He et al., 22 Sep 2025).

This interface distinguishes Chat-CBM from the earlier interactive CBM policy framework. In the earlier setting, interaction consists of selectively acquiring ground-truth concept labels according to an acquisition score such as $g$2, where $g$3 is concept uncertainty and $g$4 is concept influence on the label predictor. In Chat-CBM, by contrast, interventions can be numerical or linguistic and can include external knowledge or reasoning advice, not just correction of existing concept coordinates (Chauhan et al., 2022).

4. Training and inference pipeline

Chat-CBM preserves the underlying CBM training regime and leaves the LLM frozen. For supervised CBMs with ground-truth concept labels, the concept predictor $g$5 is trained with

$g$6

and the linear label head $g$7 is trained independently with

$g$8

The reported implementation uses AdamW and CosineAnnealingLR. For unsupervised CBMs, the bottleneck is constructed with

$g$9

for each concept $\hat c = g(x)$0 in a large CLIP-derived bank, and only $\hat c = g(x)$1 is trained on $\hat c = g(x)$2. In all cases, the LLM classifier is fully frozen, with prompts $\hat c = g(x)$3, $\hat c = g(x)$4, $\hat c = g(x)$5, and $\hat c = g(x)$6 designed by hand rather than learned (He et al., 22 Sep 2025).

The inference-time workflow is explicitly specified in seven steps. The model first computes $\hat c = g(x)$7, extracts top-$\hat c = g(x)$8 concept semantics $\hat c = g(x)$9, and generates candidate labels $g_j(x) \in [0,1]$0. It then selects $g_j(x) \in [0,1]$1 in-context examples per candidate to form $g_j(x) \in [0,1]$2, optionally computes a per-class prior $g_j(x) \in [0,1]$3 from training-set statistics given by average concept frequencies $g_j(x) \in [0,1]$4, and forms a single long prompt containing the instruction, ICL examples, class-prior sentences, and test input semantics. The LLM is called with do_sample=true, top_k=10, and max_length up to $g_j(x) \in [0,1]$5–$g_j(x) \in [0,1]$6K tokens, after which the model parses structured output of the form <analysis: …><answer: class> (He et al., 22 Sep 2025).

Ablation results indicate that prompt construction is not incidental. The paper reports that adding class-prior $g_j(x) \in [0,1]$7 and clean ICL demonstrations $g_j(x) \in [0,1]$8 each yields substantial gains, exemplified on CUB by an accuracy change from $g_j(x) \in [0,1]$9 depending on combination. It also reports that increasing the number of ICL shots improves accuracy, with an example on CUB where $j$0 candidates and $j$1 improves $j$2. Larger models such as LLaMA3-70B relative to 8B and Qwen2.5-14B relative to 7B consistently outperform smaller ones, although gains saturate beyond approximately $j$3B for some datasets (He et al., 22 Sep 2025).

5. Empirical performance

The experimental evaluation spans nine datasets divided into supervised and unsupervised CBM settings. The supervised datasets are CUB with $j$4 bird concepts and $j$5 classes, AwA2 with $j$6 animal attributes and $j$7 classes, and PBC with white blood cell attributes and $j$8 classes. The unsupervised datasets are DTD, Food-101, Flower-102, CIFAR10, CIFAR100, and ImageNet. The main metric is top-1 classification accuracy, with concept-prediction accuracy also reported where available. Supervised baselines are Hard CBM, ProbCBM, CEM, and ECBM; unsupervised baselines are LaBo and V2C-CBM. The LLMs evaluated are LLaMA3-8B/70B and Qwen2.5-7B/14B/32B/72B, all in instruct variants (He et al., 22 Sep 2025).

On supervised datasets without intervention, the paper reports the following changes from a standard CBM baseline to Chat-CBM: CUB $j$9 for a gain of $h$0 percentage points, AwA2 $h$1 for $h$2 percentage points, and PBC $h$3, described as a slight change but with substantially greater robustness under intervention. On unsupervised datasets in the $h$4-shot setting, Chat-CBM averages $h$5 percentage points above LaBo and $h$6 percentage points above V2C-CBM (He et al., 22 Sep 2025).

The intervention results are central to the model’s identity. For supervised CBMs, Chat-CBM’s accuracy is reported to quickly approach $h$7 as users correct concepts, whereas standard CBMs plateau much lower. For unsupervised CBMs, the paper states that within $h$8 natural-language interventions, Chat-CBM surpasses the all-shot baseline accuracy. These observations are consistent with the paper’s claim that semantic reasoning makes interventions effective even when raw concept activations are noisy and densely activated (He et al., 22 Sep 2025).

Setting	Reported result
CUB	CBM $h$9 $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$0 Chat-CBM $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$1
AwA2	CBM $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$2 $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$3 Chat-CBM $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$4
PBC	CBM $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$5 $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$6 Chat-CBM $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$7
Unsupervised, 2-shot	$$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$8 pp over LaBo; $$p(y=i \mid \hat c) = \mathrm{softmax}(W\hat c + b)_i,\qquad \hat y = \arg\max_i p(y=i \mid \hat c).$$9 pp over V2C-CBM
Supervised intervention	Accuracy quickly approaches $S\subset\{1,\dots,m\}$0
Unsupervised intervention	Within $S\subset\{1,\dots,m\}$1 natural-language interventions, surpasses all-shot baseline

These results should be interpreted with the model’s evaluation design in mind. Candidate labels are generated by the original head $S\subset\{1,\dots,m\}$2, and the LLM reasons over a restricted top-$S\subset\{1,\dots,m\}$3 class set rather than the full label space. A plausible implication is that Chat-CBM leverages semantic disambiguation within a shortlist rather than replacing the original CBM pipeline end-to-end.

6. Use cases, limitations, and interpretability claims

The paper identifies four use cases. In medical imaging, exemplified by PBC, experts can correct or add pathological attributes in real time. In fine-grained ecology, such as bird or flower identification, field biologists can inject domain knowledge such as “this species often has a crest.” In interactive education, students can query and refine model reasoning through natural-language feedback. In low-label domains, unsupervised CBMs combined with Chat-CBM support meaningful interventions even without concept supervision (He et al., 22 Sep 2025).

The same paper also states four limitations. First, computational cost remains materially higher than for standard CBMs because each image requires an LLM call, quantified as approximately $S\subset\{1,\dots,m\}$4 seconds per image with an $S\subset\{1,\dots,m\}$5B model on an L40. Second, incomplete concept sets remain a bottleneck: if human-defined concepts omit key distinguishing features, Chat-CBM cannot learn new latent concepts without external text prompts. Third, frozen LLMs may inject harmful or biased prior knowledge, so deployment in high-stakes settings demands alignment, filtering, or unlearning of undesirable behaviors. Fourth, long chains of intervention increase context length, potentially requiring larger GPUs or dynamic retrieval of demonstrations (He et al., 22 Sep 2025).

These constraints qualify Chat-CBM’s interpretability claims. The model preserves the CBM property that users can see and edit concept-level explanations, but the final classifier is no longer a transparent linear map over concept scores. Instead, interpretability is mediated through concept semantics, candidate labels, prompt design, and the reasoning behavior of a frozen LLM. This suggests a shift from parametric transparency to interactive, language-mediated interpretability: the bottleneck remains human-readable, yet the mapping from bottleneck to label becomes more expressive, more adaptable to intervention, and correspondingly more dependent on prompt engineering and LLM behavior.