acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Concept Bottleneck Large Language Models for Controllable Generation

Abstract

Concept Bottleneck Models (CBMs) provide an interpretable interface for inspecting and intervening on model predictions through human-understandable concepts. For language generation, a key challenge is to support control at different levels of semantic granularity while ensuring that interventions on concept representations effectively influence generated text. To address this challenge, we introduce Hierarchical Concept Bottleneck LLMs (H-CBLM), a framework for controlling language generation through coarse-grained and fine-grained concepts. H-CBLM uses coarse-grained concept logits to gate fine-grained concept logits according to a predefined hierarchy, while retaining residual features to support language modeling. We further introduce an intervention reconstruction objective that updates only the concept-to-token weights to reconstruct text after label-based concept replacement, complementing residual regularization adapted to partially observed multi-label supervision. At inference time, users can intervene on fine-grained concepts representing specific attributes or coarse-grained concepts representing broader semantic categories. Compared with the adapted CB-LLM baseline, H-CBLM increases the success rate of generation satisfying both target-concept and quality criteria by an average of 32.74 percentage points across six datasets, while coarse-grained interventions yield more evenly distributed expression of associated fine-grained concepts across all four evaluated datasets. In an LLM agent application, local coarse-grained interventions reduce tool-call attempts by 46.96%, at the cost of a minor 2.33-percentage-point decrease in task accuracy.

Primary area:
interpretability and explainable AI
Keywords:
Concept Bottleneck Models, LLM Interpretability, Controllable Text Generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_9vdMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

The authors present H-CBLM, an extension to CB-LLM that adds a hierarchical gating mechanism to enable coarse- and fine-grained control of text generation. Experiments show substantial improvements in joint steering success across six datasets.

Critical strengths
  • The paper tackles a clear problem, and adding coarse-grained control on top of fine-grained concepts is a sound idea.
  • It covers a wide variety of datasets, and the metrics (especially joint success) are well chosen.
  • The paper is generally easy to read and follow.
Critical weaknesses
  • In the main manuscript results (Tables 2 and 3), the authors compare H-CBLM against a baseline that has neither the gating nor the steering (concept replacement) loss. Since the method is presented as a hierarchical framework, it is unclear how much of the improvement is due to the hierarchy versus the loss objective. I think a flat model trained with the steering loss would be needed to isolate the contribution of the hierarchy.

  • Section 4.4 does not compare zeroing a coarse concept against a flat model where the corresponding fine-grained concepts are zeroed directly, so it is unclear what the hierarchy adds for suppression.

  • (Minor) Some appendices are not referenced in the main text, so readers must go through them one by one to find what they contain. It would be good to have pointers in the main text to the relevant sections.

Questions for authors
  • Do the gains in Tables 2 and 3 come from the loss differences with respect to CB-LLM, or do they stem from the hierarchical addition?

  • How do the authors measure the advantage of zeroing a coarse-grained concept against directly setting its associated fine-grained concepts to zero in a flat model? In which circumstances do the authors envision coarse-grained suppression being preferable, and could they provide an experiment that shows it?

  • How sensitive is H-CBLM to the value tau set for the steering loss? Does the best inference value track the training value, or one should use a different one?

4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Q1: Do the gains come from the loss or from the hierarchy?

Thanks for raising this question. To complete the 2×22\times2 design, we trained a flat CB-LLM with the same reconstruction objective (τ=100\tau=100, same data, budget, seeds and protocol). Mean ±\pm sd over 3 seeds, judged by GPT-5-mini as in the paper.

Fine-grained control (Table 2 protocol)

DatasetGatingRecon.Macro-F1HitFluencyJS
CEBaB✗✗72.48±0.2140.42±11.343.15±0.4711.67±8.51
✗✓72.59±0.6961.67±6.884.55±0.0155.83±5.77
✓✗81.64±0.4829.58±6.883.11±0.4611.67±4.73
✓✓81.02±0.2060.00±3.314.56±0.0852.92±3.15
GoEmotions✗✗3.50±1.3319.05±5.202.86±0.574.29±2.79
✗✓7.72±3.2455.48±1.974.34±0.0150.95±3.12
✓✗10.80±4.5621.67±2.533.45±0.3311.07±4.39
✓✓11.41±2.3255.00±0.714.31±0.0250.24±0.21
DBpedia✗✗98.83±0.1083.33±6.641.88±0.150.48±0.41
✗✓98.92±0.0551.90±3.383.81±0.0844.29±5.00
✓✗87.80±3.5878.33±3.601.90±0.070.24±0.41
✓✓89.86±2.5154.29±5.853.66±0.0341.43±3.78

Coarse-grained control (Table 3 protocol)

DatasetGatingRecon.Coarse HitFluencyCoarse JSFine HitFine CoverageH_norm
CEBaB✗✗62.50±12.054.05±0.0843.33±10.4160.42±13.0191.67±7.220.771±0.121
✗✓70.00±11.114.33±0.2359.17±13.7167.50±13.52100.00±0.000.939±0.009
✓✗58.33±8.514.25±0.1542.08±5.9156.25±7.8191.67±14.430.824±0.081
✓✓69.17±8.044.48±0.1562.08±7.3265.00±7.50100.00±0.000.874±0.121
GoEmotions✗✗57.92±7.643.37±0.4923.75±7.5057.92±7.6465.85±7.660.723±0.094
✗✓60.00±3.753.78±0.0842.50±5.0060.00±3.7567.74±2.520.820±0.070
✓✗57.50±8.753.44±0.4636.25±3.3157.50±8.7560.35±1.260.790±0.018
✓✓60.83±2.893.81±0.1545.42±6.1760.83±2.8969.89±4.510.847±0.036
DBpedia✗✗89.72±2.932.38±0.083.06±0.4886.94±1.92100.00±0.000.705±0.221
✗✓66.67±2.202.89±0.2212.50±7.4155.83±2.2198.15±3.210.880±0.018
✓✗83.61±7.092.53±0.205.00±3.0078.33±9.4698.15±3.210.768±0.175
✓✓66.67±0.832.87±0.1510.28±4.8854.44±2.5598.15±3.210.880±0.022

The steering gains come mainly from the reconstruction objective. It raises fine-grained JS from 0.2–11.7% to 41–56% with or without gating. With reconstruction, gating changes fine- and coarse-grained JS by at most 3 points, within seed variation.

The hierarchy instead provides category-level switching: a child is exactly zero at every token where its parent is inactive, so a single coarse coordinate switches a whole category on or off, which a flat model does not guarantee (see our response to Q2).

AuthorsOpenReview ↗

Rebuttal (2/2)

Q2: Coarse-grained suppression vs. zeroing the children in a flat model

We ran this comparison in the tool-use setting of Section 4.4, where suppression matters because the tool-call concept is active at the decision points. We trained a flat bottleneck with the same concepts (2 coarse, 15 fine tool families) and the same objectives, and intervened at the same natural tool-call positions on all 2,700 When2Tool tasks. Mean ± sd over 3 seeds; changes relative to natural decoding.

ModelInterventionΔ Accuracy (pts)Δ Tool calls / task
Flatall 15 fine concepts → 00.00±0.000.00±0.00
Flatcoarse: non-tool-call → 100, tool-call → 0−2.57±0.73−0.41±0.04
H-CBLMcoarse: non-tool-call → 100, tool-call → 0 (gate closes the tool children)−1.28±1.15−0.45±0.08

Natural accuracy is 62.3% (flat) and 62.6% (H-CBLM).

Zeroing the children does not suppress the category. In the flat model, the fine activations were positive at 83% of the 665 intervened positions, yet zeroing them changed none of the next-token decisions. The decision is carried by the coarse coordinate, which also feeds the vocabulary head directly. Suppression therefore has to act on the category. H-CBLM does this with one coordinate and guarantees that every child is closed along with it. It reduces tool calls as much as the flat coarse edit, at a comparable or lower accuracy cost.

Zeroing the parent also overrides fine-grained edits. The hierarchy matters most when a coarse switch must override fine-grained activations that are active, for example because they were set explicitly. We set a fine concept to τ (pre-gate) while zeroing its parent, or while activating a contradicting parent instead (food → service in CEBaB; positive → negative in GoEmotions), and compared with the flat model trained with the same objectives (DeepSeek-V4.1-Flash, 3 seeds; %, lower is better):

CEBaB Flat + rec.CEBaB H-CBLMGoEmotions Flat + rec.GoEmotions H-CBLM
Parent zeroed: fine concept still expressed48.021.233.09.6
Contradicting parent: fine concept and that parent both expressed30.810.89.63.4

In H-CBLM, the zeroed parent removes the fine concept to close to its natural rate, whereas the flat model still expresses it in about half (CEBaB) or a third (GoEmotions) of outputs. A GPT-5-mini re-scoring of a 40-pair subset agrees (20.8% vs. 53.3% on CEBaB; 14.2% vs. 37.5% on GoEmotions).

When coarse suppression is preferable: when the target is a category with many members, or members not known in advance (e.g., "no tool calls" rather than enumerating 15 tool families); when the category is actually active at the point of intervention; and when fine-grained activations may contradict the coarse decision, as above. By contrast, on CEBaB and GoEmotions with neutral prompts, the target concepts are rarely active. Zeroing the parent, the children, or both changes concept presence by less than 2 points for every model, so zeroing has little to suppress there.

Q3: Sensitivity to the training τ, and whether the best inference value tracks it

We retrained H-CBLM on CEBaB with training τ ∈ {50, 100, 200} (3 seeds each, identical pipeline otherwise) and swept the inference value over {10, 50, 100, 150, 200, 300, 500} with the protocol of Table C.2. Fine-grained JS (%), mean ± sd over 3 seeds; the best inference value per row is in bold.

Training τ \ Inference1050100150200300500
5016.7±5.142.9±5.143.3±2.613.8±4.32.1±0.70.0±0.00.0±0.0
10015.0±2.228.3±2.938.8±2.236.7±6.216.2±3.31.2±1.20.0±0.0
20019.6±4.723.8±4.532.1±4.742.1±3.837.2±3.221.7±10.50.8±0.7

The best inference value tracks the training τ. The JS peak moves from 50–100 (τ = 50) to 100 (τ = 100) to 150–200 (τ = 200). Hit and fluency shift in the same way: Hit peaks at 150, 200 and 300, and fluency falls below 3 at 150, 200 and 300 for τ = 50, 100 and 200, respectively. The peak JS itself is insensitive to τ (38.8–43.3%). The best inference value is at or slightly above the training value; with the 100-token cap of Tables C.2–C.4, JS for τ = 100 models also peaks at 150–200. We nevertheless use the training value (τ = 100) as the intervention value in all main results, fixed in advance without any selection on test outputs.

To control cost, these runs were scored with DeepSeek-V4.1-Flash instead of GPT-5-mini. Absolute scores may differ between evaluators, but the trends we report are robust to this choice.

Minor: Appendix references

Thank you for pointing this out. In the revision, we will add pointers in the main text to every appendix section where the related content is discussed, so that readers can locate the supporting details directly.

Thank you for running the additional experiments and addressing my questions.

  • I appreciate the new experiments comparing coarse-grained suppression with directly zeroing the fine-grained concepts in a flat model. The experiments involving conflicting parent and child interventions help demonstrate the practical utility of hierarchical gating.

  • The sensitivity analysis on τ\tau is also useful. I think it would be useful to translate this into a practical recommendation for users in the paper.

  • My main remaining concern is that the new ablations suggest that the improvements reported in the main results come primarily from the reconstruction objective rather than the hierarchical gating mechanism. I acknowledge that the hierarchy provides additional value, as demonstrated by the new experiments in Q2. However, I believe the current manuscript does not sufficiently disentangle these contributions or adequately reflect their respective roles in the reported improvements.

I appreciate the additional work, but I believe that the findings would require a broader revision of the manuscript, particularly in how the contributions are framed and how the main results are presented and interpreted. I therefore maintain my original score.

Thank you for your valuable feedback and suggestions. We agree that the manuscript did not sufficiently disentangle the contributions of the reconstruction objective and the hierarchical gating. We have revised the framing of the contributions in the abstract, introduction, Section 4.3, and Appendix C.1. The revised version now states that the improvements in the main results come primarily from the reconstruction objective, while the hierarchical gating provides complementary benefits for controllability, as shown by the coarse-grained suppression and conflicting parent–child intervention experiments. However, the analysis of the hierarchical design still needs more evidence to support it. We fully respect your assessment. Thank you again for your time and constructive comments.

Reviewer_GhKAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper introduces H-CBLM, which proposes coarse-to-fine concept gating and an intervention reconstruction objective for controllable language generation. Reconstruction trains concept-to-token weights under edited concept activations. Experiments compare the method with an adapted flat CB-LLM on six datasets and demonstrate local tool-call suppression.

Critical strengths

The intervention reconstruction addresses a concrete training-inference mismatch; Table C.1 retains the hierarchy while removing reconstruction. Joint semantic-and-fluency success falls from 52.92% to 11.67% on CEBaB and from 41.43% to 0.24% on DBpedia. This is strong evidence for the reconstruction objective.

Critical weaknesses

As in Eq (5), the coarse concepts can directly influence token logits, so improved coarse control could arise through this path without hierarchical gating, so the hierarchy's incremental contribution is unresolved. The main comparison changes both gating and reconstruction, while Table C.1 removes only reconstruction. A flat model with reconstruction is missing (Appendix A.4). The existing strength sweeps do not isolate these effects.

Questions for authors

Can you add a flat model trained with the same intervention reconstruction objective as H-CBLM, keeping the other training and evaluation settings matched? Together with the existing models, this would complete the comparison of gating and reconstruction, each enabled or disabled. Please report both fine- and coarse-grained control, including joint success, fluency, and fine-concept coverage.

Since coarse concepts directly affect token logits in Eq. (5), how much does updating the fine-concept gates add to coarse control? Using the same trained H-CBLM, can you compare a coarse intervention that keeps fine-concept activations at their pre-intervention values with the standard intervention that also updates them through gating? Comparing these conditions across intervention strengths would help separate the direct coarse-to-token effect from the effect through gating.

4 comments
AuthorsOpenReview ↗

Rebuttal (1/2)

Q1: Flat model with reconstruction (gating × reconstruction)

Thank you for this suggestion. We trained a flat CB-LLM with the same reconstruction objective (τ=100\tau=100, same data, budget, seeds and protocol). Mean ±\pm sd over 3 seeds, judged by GPT-5-mini as in the paper.

Fine-grained control (Table 2 protocol)

DatasetGatingRecon.Macro-F1HitFluencyJS
CEBaB✗✗72.48±0.2140.42±11.343.15±0.4711.67±8.51
✗✓72.59±0.6961.67±6.884.55±0.0155.83±5.77
✓✗81.64±0.4829.58±6.883.11±0.4611.67±4.73
✓✓81.02±0.2060.00±3.314.56±0.0852.92±3.15
GoEmotions✗✗3.50±1.3319.05±5.202.86±0.574.29±2.79
✗✓7.72±3.2455.48±1.974.34±0.0150.95±3.12
✓✗10.80±4.5621.67±2.533.45±0.3311.07±4.39
✓✓11.41±2.3255.00±0.714.31±0.0250.24±0.21
DBpedia✗✗98.83±0.1083.33±6.641.88±0.150.48±0.41
✗✓98.92±0.0551.90±3.383.81±0.0844.29±5.00
✓✗87.80±3.5878.33±3.601.90±0.070.24±0.41
✓✓89.86±2.5154.29±5.853.66±0.0341.43±3.78

Coarse-grained control (Table 3 protocol)

DatasetGatingRecon.Coarse HitFluencyCoarse JSFine HitFine CoverageH_norm
CEBaB✗✗62.50±12.054.05±0.0843.33±10.4160.42±13.0191.67±7.220.771±0.121
✗✓70.00±11.114.33±0.2359.17±13.7167.50±13.52100.00±0.000.939±0.009
✓✗58.33±8.514.25±0.1542.08±5.9156.25±7.8191.67±14.430.824±0.081
✓✓69.17±8.044.48±0.1562.08±7.3265.00±7.50100.00±0.000.874±0.121
GoEmotions✗✗57.92±7.643.37±0.4923.75±7.5057.92±7.6465.85±7.660.723±0.094
✗✓60.00±3.753.78±0.0842.50±5.0060.00±3.7567.74±2.520.820±0.070
✓✗57.50±8.753.44±0.4636.25±3.3157.50±8.7560.35±1.260.790±0.018
✓✓60.83±2.893.81±0.1545.42±6.1760.83±2.8969.89±4.510.847±0.036
DBpedia✗✗89.72±2.932.38±0.083.06±0.4886.94±1.92100.00±0.000.705±0.221
✗✓66.67±2.202.89±0.2212.50±7.4155.83±2.2198.15±3.210.880±0.018
✓✗83.61±7.092.53±0.205.00±3.0078.33±9.4698.15±3.210.768±0.175
✓✓66.67±0.832.87±0.1510.28±4.8854.44±2.5598.15±3.210.880±0.022

The steering gains come mainly from the reconstruction objective. It raises fine-grained JS from 0.2–11.7% to 41–56% with or without gating. With reconstruction, gating changes fine- and coarse-grained JS by at most 3 points, within seed variation.

The hierarchy instead provides category-level switching: a child is exactly zero at every token where its parent is inactive, so a single coarse coordinate switches a whole category on or off, which a flat model does not guarantee

AuthorsOpenReview ↗

Rebuttal (2/2)

Q2: How much does updating the fine-concept gates add to coarse control?

We ran exactly this comparison. Using the H-CBLM checkpoints from our main experiments (3 seeds), we raised the target coarse concept to each strength in {10, 50, 100, 150, 200, 300, 500} under two conditions. Gated is the standard intervention, in which the children are recomputed through the gate. Frozen keeps every fine-concept activation at its pre-intervention (post-gate) value. All other settings follow Table C.2. Coarse Hit and Coarse JS (%), seed means, scored with DeepSeek-V4.1-Flash on all outputs.

DatasetStrengthHit (frozen)Hit (gated)JS (frozen)JS (gated)
CEBaB1052.151.744.644.2
5057.958.344.243.8
10066.265.840.440.8
15077.578.320.420.8
20081.581.56.76.7
30051.451.80.80.8
5009.69.20.00.0
GoEmotions1036.336.219.220.4
5044.444.315.617.7
10057.959.710.911.7
15056.257.31.21.9
20048.848.70.00.0
30012.914.70.00.0
50013.014.10.00.0

Updating the gates adds little to coarse control. At every strength, the two conditions differ by at most 1.8 points in Hit and 2.1 points in JS, and the generated texts are often identical (95–99% on CEBaB, 62–80% on GoEmotions).

Two properties explain this. First, the gate g = 1[z > 0]·σ(z) takes values only in {0} ∪ [0.5, 1), so updating it can at most double a child that is already active, and opening an inactive parent lets a child reach at most its raw value (negative raw values remain zero after the ReLU in the head). Second, natural fine-concept activations are small: on 1,000 held-out validation texts, the document-averaged effective fine activations of the same checkpoints are positive in 27–28% of cases on CEBaB, with a median of about 1.5 and a 99th percentile below 7. Any change through the gate is therefore small next to a coarse concept set to 50–200. Consistent with this, the gated children are practically identical from strength 10 to 200 (σ(10) > 0.9999), yet CEBaB Hit rises from about 52 to 81.5 under both conditions. The effect of strength thus comes from the value of the coarse concept itself, through its own term in the token logits (Eq. (5)).

The gate's role is therefore structural rather than a source of control strength. It keeps fine-grained activations consistent with the coarse decision. To show this, we edited a fine concept against its parent: the fine concept is set to τ (pre-gate) while its parent is zeroed, or while a contradicting parent is activated instead (food → service in CEBaB; positive → negative in GoEmotions). We compare H-CBLM with the flat model trained with the same reconstruction objective (DeepSeek-V4.1-Flash, 3 seeds; %, lower is better):

CEBaB Flat + rec.CEBaB H-CBLMGoEmotions Flat + rec.GoEmotions H-CBLM
Parent zeroed: fine concept still expressed48.021.233.09.6
Contradicting parent: fine concept and that parent both expressed30.810.89.63.4

In H-CBLM, the coarse decision overrides the inconsistent fine edit. The remaining rate is close to the concept's natural rate in unedited text, and self-contradictory outputs are about 3× less frequent than in the flat model, which has no mechanism to enforce this consistency. A GPT-5-mini re-scoring of a 40-pair subset gives the same picture (parent zeroed: 20.8% vs. 53.3% on CEBaB, 14.2% vs. 37.5% on GoEmotions).

Thank you for running the requested experiments and reporting the results clearly.

The new ablations show that reconstruction accounts for most of the steering gains. With reconstruction matched, the flat model performs similarly to H-CBLM, and the frozen-versus-gated comparison shows little benefit from updating the fine-concept gates during positive coarse control. These results weaken the case for hierarchical gating as a central contribution.

The conflict tests show that gating reduces inconsistent outputs under conflicting parent–child edits. However, this benefit follows a predefined rule that gives parent concepts priority over their children. I find this contribution limited, particularly given the lack of clear gains in the standard steering experiments.

I continue to see value in the reconstruction objective, but the hierarchy’s added value is insufficient to change my overall assessment. I therefore maintain my original score.

Thank you for your careful evaluation and constructive feedback. We agree that the reconstruction objective accounts for most of the steering gains. Regarding the conflict tests, the parent-priority rule is intended as a structural guarantee that a flat model does not provide, and we will state this design choice and its scope more explicitly. We respect your assessment and thank you again for your time.

Reviewer_yoYnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces H-CBLM, a concept bottleneck language model with coarse-to-fine gating and an intervention reconstruction objective. Coarse concepts gate fine concepts, while residual features support next-token prediction. Reconstruction training updates the concept-to-token weights under label-based activation replacement. Across six datasets, the complete model increases average joint concept-hit and fluency success from 7.63% to 40.37% relative to an adapted CB-LLM. Coarse interventions increase mean concept-distribution entropy on four datasets. An exploratory When2Tool experiment reduces tool-call attempts by 46.96%, with a 2.33-percentage-point accuracy decrease.

Critical strengths
  1. The architecture provides a clear interface for coordinated intervention. The gating mechanism explicitly connects a parent concept to its children, and the training procedure specifies which parameters receive reconstruction gradients. This is a useful design for controlling semantic categories at different granularities, with a training objective directly addressing the gap between concept recognition and usable generation control.
  2. The paper evaluates control together with generation quality. The experiments report concept hit, fluency, joint success, and semantic breadth across three training seeds; the reconstruction ablation shows substantial reconstruction gains even with gating retained. The agent experiment also reports the accuracy cost of reducing calls, making the observed intervention trade-off explicit.
Critical weaknesses
  1. The evidence does not isolate the contribution of hierarchy. The baseline configuration removes both gating and reconstruction from CB-LLM, while the ablation removes only reconstruction from H-CBLM. A flat bottleneck trained with the same reconstruction objective is therefore essential. Without this control, the large gains cannot substantiate the specific value of hierarchical gating over training the vocabulary head for edited activations.
  2. The generation evaluation provides limited independent support for the broad control claim. The generation evaluation uses one fixed instruction per text dataset, only 49 unique GSM8K questions, and ten ToxicChat questions. GPT-5-mini both supplies weak QA annotations and judges generation. Prompt-diverse evaluation and an independently assessed output subset are needed to establish that the gains extend beyond this narrow evaluation setup.
Questions for authors
  1. Do you have a flat CB-LLM trained with the same intervention reconstruction objective, label treatment, and budget? Its fine- and coarse-control results would distinguish the contribution of hierarchy from reconstruction training.
  2. Were intervention strengths selected on a separate validation set for each method? The intervention-strength sweeps show large changes in baseline fluency as strength changes. Please identify the selection protocol and the comparison at independently selected strengths.
  3. Were the five GSM8K steering targets and the test question sets fixed before test-output inspection, and is there any independent human or second-judge assessment of the generated outputs? Please specify the selection split and available agreement evidence.
3 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

Q1: Flat CB-LLM with the same reconstruction objective

Yes. We trained a flat CB-LLM with the same reconstruction objective (τ=100\tau=100, same data, label treatment, budget, seeds and protocol). Mean ±\pm sd over 3 seeds, judged by GPT-5-mini as in the paper. Fine-grained control (Table 2 protocol)

DatasetGatingRecon.Macro-F1HitFluencyJS
CEBaB✗✗72.48±0.2140.42±11.343.15±0.4711.67±8.51
✗✓72.59±0.6961.67±6.884.55±0.0155.83±5.77
✓✗81.64±0.4829.58±6.883.11±0.4611.67±4.73
✓✓81.02±0.2060.00±3.314.56±0.0852.92±3.15
GoEmotions✗✗3.50±1.3319.05±5.202.86±0.574.29±2.79
✗✓7.72±3.2455.48±1.974.34±0.0150.95±3.12
✓✗10.80±4.5621.67±2.533.45±0.3311.07±4.39
✓✓11.41±2.3255.00±0.714.31±0.0250.24±0.21
DBpedia✗✗98.83±0.1083.33±6.641.88±0.150.48±0.41
✗✓98.92±0.0551.90±3.383.81±0.0844.29±5.00
✓✗87.80±3.5878.33±3.601.90±0.070.24±0.41
✓✓89.86±2.5154.29±5.853.66±0.0341.43±3.78

Coarse-grained control (Table 3 protocol)

DatasetGatingRecon.Coarse HitFluencyCoarse JSFine HitFine CoverageH_norm
CEBaB✗✗62.50±12.054.05±0.0843.33±10.4160.42±13.0191.67±7.220.771±0.121
✗✓70.00±11.114.33±0.2359.17±13.7167.50±13.52100.00±0.000.939±0.009
✓✗58.33±8.514.25±0.1542.08±5.9156.25±7.8191.67±14.430.824±0.081
✓✓69.17±8.044.48±0.1562.08±7.3265.00±7.50100.00±0.000.874±0.121
GoEmotions✗✗57.92±7.643.37±0.4923.75±7.5057.92±7.6465.85±7.660.723±0.094
✗✓60.00±3.753.78±0.0842.50±5.0060.00±3.7567.74±2.520.820±0.070
✓✗57.50±8.753.44±0.4636.25±3.3157.50±8.7560.35±1.260.790±0.018
✓✓60.83±2.893.81±0.1545.42±6.1760.83±2.8969.89±4.510.847±0.036
DBpedia✗✗89.72±2.932.38±0.083.06±0.4886.94±1.92100.00±0.000.705±0.221
✗✓66.67±2.202.89±0.2212.50±7.4155.83±2.2198.15±3.210.880±0.018
✓✗83.61±7.092.53±0.205.00±3.0078.33±9.4698.15±3.210.768±0.175
✓✓66.67±0.832.87±0.1510.28±4.8854.44±2.5598.15±3.210.880±0.022

The steering gains come mainly from the reconstruction objective. It raises fine-grained JS from 0.2–11.7% to 41–56% with or without gating. With reconstruction, gating changes fine- and coarse-grained JS by at most 3 points, within seed variation.

The hierarchy instead provides consistency between the two levels. A child is exactly zero at every token where its parent is inactive, so a single coarse coordinate switches a whole category off and overrides contradicting fine-grained activations, which a flat model does not guarantee. To test this, we edited a fine concept against its parent: the fine concept is set to τ (pre-gate) while its parent is zeroed, or while a contradicting parent is activated instead (food → service in CEBaB; positive → negative in GoEmotions). We compare H-CBLM with the flat model trained with the same reconstruction objective (DeepSeek-V4.1-Flash as judge, 3 seeds; %, lower is better):

CEBaB Flat + rec.CEBaB H-CBLMGoEmotions Flat + rec.GoEmotions H-CBLM
Parent zeroed: fine concept still expressed48.021.233.09.6
Contradicting parent: fine concept and that parent both expressed30.810.89.63.4

In H-CBLM, the coarse decision overrides the inconsistent fine edit. The remaining rate is close to the concept's natural rate in unedited text, and self-contradictory outputs are about 3× less frequent than in the flat model, which has no mechanism to enforce this consistency. A GPT-5-mini re-scoring of a 40-pair subset gives the same picture (parent zeroed: 20.8% vs. 53.3% on CEBaB, 14.2% vs. 37.5% on GoEmotions).

AuthorsOpenReview ↗

Rebuttal (2/3)

Q2: Selection of intervention strengths

No strength was selected on test outputs. Every method uses the training value τ = 100 as its intervention value, fixed in advance. To check that this choice does not disadvantage any method, we compare each method at its best strength in the sweeps of Tables C.2–C.4. These are selected on the test outputs, which favors the baseline. Fine-grained JS (%); the sweeps use a 100-token cap, so the τ = 100 values differ slightly from Table 2:

DatasetCB-LLM, best strengthH-CBLM at τ = 100H-CBLM, best strength
CEBaB25.00±5.73 (50)54.58±5.9164.58±7.32 (150)
GoEmotions13.33±1.35 (natural)51.79±1.2959.17±3.00 (150)
DBpedia18.10±1.49 (50)37.38±5.0759.76±6.07 (200)

Even against the baseline's best strength, H-CBLM at the fixed training value is 2–4× higher in JS. The large fluency changes in the sweeps occur mainly for CB-LLM, which was not trained for edited activations. We also retrained H-CBLM on CEBaB with τ ∈ {50, 100, 200}: the best inference value moves with the training τ (JS peaks at 50–100, 100 and 150–200, respectively), so the training value is a principled default that requires no test-time tuning.

AuthorsOpenReview ↗

Rebuttal (3/3)

Q3: GSM8K targets, evaluation breadth, and independent assessment (also Weakness 2)

Selection. The five targets and the questions were fixed before any test outputs were generated; the targets were chosen by a brief semantic screening of the concept inventory, without using test-set outputs. In QA, which operations appear depends mainly on the question, so we chose general operations that can enter many solutions without changing the answer (addition, subtraction, multiplication, division, linear equation; 5 of 14 concepts). The others (e.g., percent change, profit and loss) are tied to specific problem types. Questions come from the test split by a fixed rule: per target, 5 whose reference solution uses the operation and 5 whose does not.

Larger evaluation set. With 50 GSM8K questions per concept (250 per model and seed) and 100 ToxicChat questions, results match the original subsets (GPT-5-mini, 3 seeds; original subsets in parentheses):

DatasetModelHit: extended (original)JS: extended (original)
GSM8KCB-LLM26.0 (26.7)3.1 (2.7)
H-CBLM48.0 (47.3)30.8 (30.7)
ToxicChatCB-LLM94.3 (93.3)36.7 (26.7)
H-CBLM94.3 (86.7)57.7 (46.7)

Independent judges. A judge from a different model family (DeepSeek-V4.1-Flash, same rubric) re-scored all text and QA outputs. Hit agreement with GPT-5-mini is 89.5% (κ = 0.75), and fluency agrees within one point on 91.1%. The H-CBLM vs. CB-LLM JS differences keep their sign for fine-, multi- and coarse-grained control on CEBaB and GoEmotions, fine-grained control on DBpedia, and GSM8K. A third judge (Claude Opus 5.5) scored a blinded random sample of 150 fine-grained interventions on CEBaB, GoEmotions and DBpedia and agrees with GPT-5-mini on 88.0% of hit decisions; under it, H-CBLM again clearly outperforms CB-LLM (53.3% vs. 33.3% Hit; 42.7% vs. 0.0% JS).

Prompt diversity. With three paraphrased instructions per dataset that differ from the training instruction, results stay close to the original (DeepSeek-V4.1-Flash as judge, 3 seeds; original instruction in parentheses). Fine-grained JS is 44.2 (37.9) for H-CBLM vs. 6.3 (4.6) for CB-LLM on CEBaB, and 29.5 (25.8) vs. 0.6 (0.6) on GoEmotions; coarse-grained results are similarly stable. No single instruction changes H-CBLM's Hit by more than 5 points.

Reviewer_4fFZOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a hierarchical concept bottleneck model for steerable language generation. The authors extend previous work by defining a hierarchy of coarse and fine-grained concepts that allows to improve the control of the generative process via interventions. Additionally, they propose a reconstruction loss to train the concept-to-token module, and include a residual unsupervised representation that is regularized following an adversarial training strategy.

Critical strengths

The method moves from independent concepts to a hierarchical structure that includes coarse semantic concepts and fine-grained concepts, including a gating system that allows a better control of the generative process via interventions. The gating system ensures that interventions and generation are consistent with the defined graph. I believe that modeling these structures and concept relationships is a valuable research direction.

Critical weaknesses
  • I think the paper is lacking a clear motivation beyond improving the previous existing approach for text generation using CBMs (CB-LLMs).
  • I think some of the design choices are not sufficiently justified in the text, whether intuitively, theoretically, technically, or empirically (for example, why are the concepts assumed to be binary? why is only the vocabulary regularized? why joint training instead of sequential?)
  • I think the empirical evaluation can be improved.
Questions for authors
  • Why should we control the text generation via concept interventions instead of prompting? What is the main motivation? Is there any real gain in controlling the text generation via concept intervention (either in user interaction, model auditing, or performance)?
  • When intervening the fine-grained concepts, the logit of the parent coarse concept is also intervened. Does this affect how the model process the rest of the fine-grained concepts that share the same parent?
  • For training this model, do we need to assume that we have access to ground-truth interventional data? How realistic is this in practice?
  • The authors follow a joint training approach, including a steering loss to train the vocabulary head. However, in CBMs it is common to follow a sequential training approach. First, the concept predictor is trained. Then, in a second phase, the concept predictor is frozen and the task predictor is trained using the concepts. Following such an approach would avoid the need for interventions during training, and would reduce potential concept leakage issues. Have the authors considered this? If so, what advantages does their joint training approach offer over sequential training?
  • Regarding the evaluation
    • The authors apply an adversarial training strategy to enforce disentanglement between concepts and residual. How much computational overload does this strategy introduce during training? Also, the disentanglement between concepts and the residual connection is not empirically evaluated. One possible experiment could be to modify the residual while keeping the concepts fixed, and see how this affects the content of the text generated.
    • During evaluation, do the authors assume that only one concept is intervened on at a time? Is it possible to intervene multiple concepts at once? How does the model handle concept relationships beyond hierarchy (for example, how does it handle the activation of mutually exclusive concepts)?
    • In some cases, the hit performance after concept intervention is below 50%. Do the authors have any intuition on why the model struggles to generate text aligned with the specified concept configuration in some datasets?
    • I think it would be helpful to give a bit more of context on the LLM Agent tool use case, specially for readers from the CBM community that are not familiar with LLM literature. Which type of tasks is the model solving? Is it more important to maintain accuracy or have fewer tool calls in this task? How does the baseline CB-LLM performs in this setting? When does the model ignore the provided concept configuration and calls a tool?
5 comments
AuthorsOpenReview ↗

Rebuttal (1/3)

Q1: Motivation — why concept interventions rather than prompting?

Prompting, in-context learning, and other controllable generation techniques steer the outputs of an otherwise black-box LLM. They can achieve effective control, but they do not expose explicit intermediate representations that can be inspected, measured, or intervened upon. A concept bottleneck routes the hidden state through named concept coordinates. These can be inspected as the model's predictions of which concepts the text expresses, and edited to steer what is generated. Control is therefore more transparent: an edit acts on a known internal variable rather than on the model's interpretation of an instruction, and its effect can be measured against that variable. The hierarchy organizes these coordinates as categories with more specific attributes, so the model can be inspected and controlled at either granularity. We see concept interventions as complementary to prompting, not as a replacement for it.

Q2: Does raising the parent affect the siblings?

We raise the parent together with the target fine concept to keep the edit consistent with the hierarchy. Under gating, a child can only be active when its parent is, so a fine concept at τ under an inactive parent is a configuration the model never sees in training. The side effect on the siblings is small. The gate takes values only in {0} ∪ [0.5, 1), so raising the parent can at most double an active sibling, and where the parent was inactive, it lets a sibling reach at most its raw value (negative raw values remain zero after the ReLU).

To test whether the gain in target Hit comes from the siblings, we compared the standard intervention with one that raises the parent but freezes every sibling at its pre-intervention value (3 seeds). Freezing changes target Hit by at most 2.5 points and sibling expression by less than 1 point on CEBaB and GoEmotions, so the target's Hit does not come from activating its siblings. Raising the parent does contribute directly, through its own term in the vocabulary head: compared with leaving the parent at its natural value, it raises target Hit by about 15 points.

Q3: Is ground-truth interventional data required?

No. Reconstruction (Eq. 8) uses the ordinary training texts and their concept labels. Concepts labeled positive are set to τ, those labeled negative are set to zero, and unlabeled concepts keep their natural values; the vocabulary head then learns to predict the observed next token of the same text. No counterfactual or paired text is needed, only the concept labels that any concept bottleneck model requires.

Q4: Joint vs. sequential training

We considered it, but expect two problems in the generation setting. First, fluency: in a sequential scheme the backbone (LoRA) would be trained only for concept prediction, and next-token prediction would have to be recovered afterwards by the vocabulary head alone, on representations that were not shaped for it. Joint training keeps the language-modeling loss on the backbone, so the concept and residual coordinates remain predictive of the next token. Second, interventions: a head trained only on natural activations sees an edited activation at τ out of distribution. Table C.1 shows the consequence even under joint training: without reconstruction, JS drops from 52.9% to 11.7% on CEBaB. Reconstruction updates only the concept columns of the head and leaves the concept predictor unchanged, so it could also serve as a second-stage objective. Leakage mainly arises through the residual pathway that next-token prediction requires, which exists under either scheme; we address it with the adversarial regularization and reconstruction objective. We follow CB-LLM's joint training in the generation setting.

AuthorsOpenReview ↗

Rebuttal (2/3)

Q5: Evaluation

Q5.1: Adversarial regularization — cost, role, and concept–residual interaction

Cost. The adversarial steps train a small probe on the 768-dimensional residual. On CEBaB (3 seeds, same GPU type), they slow training by about 2% (2.48 vs. 2.52 steps/s).

Role. We keep the adversarial regularization as in CB-LLM, where it was introduced to keep concept information out of the residual and was shown to improve steerability. H-CBLM additionally introduces the reconstruction objective, which directly trains the concept-to-token weights so that edited concepts yield coherent, concept-consistent text. As stated in the abstract, the two are complementary, and with reconstruction in place the adversarial loss has little additional effect: retraining H-CBLM on CEBaB without it leaves control and recognition essentially unchanged (fine-grained JS 46.2±5.7 without vs. 40.4±1.9 with it, DeepSeek-V4.1-Flash; Macro-F1 82.2 vs. 80.7). We retain it so that the comparison with CB-LLM stays controlled.

Concept–residual interaction. The output logit is the sum of a residual term and the concept term. We examined how the two divide the work by ranking each concept's vocabulary weights only over tokens the language model actually produces (≥5 occurrences in unedited outputs). The top tokens are then semantically clean, e.g., amusement: "lol, haha, funny"; gratitude: "thank, appreciate"; fear: "afraid, scared" (GoEmotions); service: "service, waited" (CEBaB). Without reconstruction (CB-LLM), the same procedure mostly yields function words (sadness: "what, because, when"). The residual therefore determines which tokens are plausible in context, and the concept redistributes probability among them toward its meaning; the semantics of an edit comes from the concept term, not from the residual.

Q5.2: Multiple concepts at once and mutually exclusive concepts

Multiple concepts. Interventions are not limited to one concept. The CEBaB evaluation already includes configurations that specify 2–4 aspects at once (e.g., good food and bad service): every target is set to τ, its parent is opened, and the opposite polarity of each specified aspect is set to zero. We report two rates over steered outputs (%): all targets present counts as a hit only if every specified target is expressed; mean fraction of targets present averages the share of specified targets that are expressed (e.g., 2 of 3 targets gives 2/3). GPT-5-mini, 3 seeds; "Flat + rec." is a flat bottleneck (CB-LLM) trained with the same objectives as H-CBLM:

# target concepts1234
All targets present (Hit)
CB-LLM41.01.90.00.0
Flat + rec.62.630.79.28.9
H-CBLM61.032.610.08.9
Mean fraction of targets present
CB-LLM41.011.92.90.0
Flat + rec.62.652.949.245.0
H-CBLM61.057.549.251.1

H-CBLM handles multi-concept configurations well. Each additional target is still expressed in about half of the outputs, while fluency stays at 4.5/5. In contrast, CB-LLM almost never expresses two or more targets, and its fluency falls to 1.3–2.0. The drop in joint Hit mainly reflects the requirement that all targets appear in one short text. H-CBLM also reduces the opposite polarity of the specified aspects relative to the clean text (53.1 → 40.2, 61.7 → 47.5, and 64.4 → 51.1 for 2, 3, and 4 targets).

Mutually exclusive concepts. The hierarchy explicitly encodes the parent–child relation, and gating enforces it. We tested contradictory configurations: a fine concept is set to τ (pre-gate) while its parent is zeroed, or while a contradicting parent is activated instead (food → service in CEBaB; positive → negative in GoEmotions). We compare H-CBLM with the flat model trained with the same concepts and objectives (DeepSeek-V4.1-Flash, 3 seeds; lower is better):

CEBaB Flat + rec.CEBaB H-CBLMGoEmotions Flat + rec.GoEmotions H-CBLM
Parent zeroed: fine concept still expressed48.021.233.09.6
Contradicting parent: fine concept and that parent both expressed30.810.89.63.4

In H-CBLM, the coarse decision overrides inconsistent fine-grained edits. The remaining rate (21.2% / 9.6%) is close to the concept's natural rate in unedited text, and self-contradictory outputs are about 3× less frequent than in the flat model. A flat bottleneck has no mechanism to guarantee this consistency. Exclusivity among siblings, such as the two polarities of an aspect, is not hard-coded. It is handled at intervention time by setting the excluded concept to zero, and it is partly learned from data, as the reduced opposite-polarity rates above show. Explicit exclusivity groups are a natural extension of the gating mechanism.

AuthorsOpenReview ↗

Rebuttal (3/3)

Q5.3: Why is Hit below 50% on some datasets?

For H-CBLM, this occurs on IMDb (fine-grained 38.9, coarse-grained 48.6) and GSM8K (47.3). We see three reasons:

  1. Fixed, untuned strength. All results use the training value τ = 100. Hit rises with strength: from τ = 100 to 200 it goes from 61 to 80 on CEBaB, 56 to 79 on GoEmotions, and 55 to 76 on DBpedia (Tables C.2–C.4), at some cost in fluency.
  2. The question constrains the answer in QA. In QA, the output depends not only on the concepts but also on the question, and a useful answer must still solve it, so steering has to work within task completion. Half of the questions for each target operation do not need it (Appendix B.2), and simple math problems often admit essentially one solution path. Steering then cannot force an operation the solution does not use without breaking the answer. Consistent with this, on questions whose reference solution uses the target operation, steering raises its use from 74.7% to 82.7% for H-CBLM, whereas on questions that do not need it, the operation rarely appears with or without steering (9.3% → 12.0%). CB-LLM instead disrupts the solution: on questions that need the operation, its use falls from 69.3% to 37.3%, and fluency drops to 2.0 (JS 30.7 vs. 2.7).
  3. Overlapping fine-grained concepts. IMDb has 26 overlapping genres, and even recognition Macro-F1 is only 44%. Outputs often express the right category but a neighboring concept: on IMDb, the target's category appears in about 69% of interventions, and on GoEmotions, 60% of misses express a sibling emotion instead (e.g., sadness for remorse).

JS, which requires both the concept and fluent text, improves over CB-LLM on all six datasets.

Q5.4: Context for the tool-use experiment

  • Tool use in LLM agents. The model is given a list of available tools in its input. When it needs one, it writes a structured call; the call is executed, its result is appended to the context, and the model continues the task from there. In H-CBLM, the coarse tool-call concept is active where the model starts such a call, which is where we intervene.
  • Tasks. When2Tool contains 2,700 test tasks at three difficulty levels. Some tasks can be answered directly, and others benefit from one of 15 tool families (e.g., calculator, code execution, retrieval). The model decides at each turn whether to call a tool.
  • Accuracy vs. tool calls. Neither is preferred in general. Tool calls cost latency and money and may expose data, so users may want fewer of them. The point of the experiment is that the concept interface turns this into an explicit, measurable trade-off: setting the non-tool-call concept to τ and the tool-call concept to zero at the natural tool-call positions reduces calls by 0.45 per task, at an accuracy cost of 1.3 points.
  • Flat baseline. We trained a flat bottleneck with the same concepts and objectives. The same category-level edit gives a similar reduction in calls (−0.41 per task) at an accuracy cost of 2.6 points, while zeroing its 15 fine-grained tool concepts changes no decision.
  • When the model still calls a tool. Calls persist mainly where the task is hard to solve without a tool (H-CBLM, 3 seeds). Among tasks that naturally call a tool, 94% of hard tasks still do, versus 68–69% of easy and medium tasks. By tool family, calls are almost unchanged for decoding (100% → 99% of tasks) and hashing (97% → 90%), which cannot be computed reliably without a tool. They largely disappear for list manipulation (15% → 1%) and code execution (28% → 14%), which the model can do directly. On hard tasks, the intervention mainly reduces the number of calls per task (2.11 → 1.26, Table 4) rather than eliminating tool use.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.