acceptodds
Under review as a conference paper at ICLR 2027

ProtoT v2: A Simpler, More Stable, and Editable Prototype Transformer

Abstract

While state-of-the-art large language models dominate benchmarks and real-world performance, their reasoning has remained hard to trust and verify. This has inspired a new generation of language models based on interpretable components, but such models often suffer from poor scalability, efficiency, and performance. In this work, we significantly simplify and stabilize the recently published Prototype Transformer (ProtoT), leading to substantial improvements across all three of these dimensions. We also develop a custom Triton kernel for the simplified formulation that increases the speed by about , substantially surpassing all baselines. Finally, we introduce a new synthetic-prototype method that enables more flexible and powerful model editing, on par with strong baselines, while showing that our model is particularly editable. Together, these advances yield ProtoT v2, a substantially more performant, efficient, and editable prototype-based language model that retains the interpretability and robustness properties of ProtoT.

Primary area:
foundation or frontier models, including LLMs
Keywords:
prototype transformer, prototype-based language models, model interpretability, nameable concept discovery, targeted model editing, steerability, linear-time sequence modeling, transformer alternatives, robustness to input perturbations, kernel development, autoregressive LMs, language models

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_BW6uOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ProtoT V2, it build upon the original architecture (ProtoT, ICML 2026) by improving its computational efficiency, numerical stability, and language modeling performance. Additionally, it introduces a synthetic-prototype-based editing mechanism to enable more flexible and targeted model interventions.

Critical strengths

This paper builds upon the original ProtoT architecture and introduces several improvements in computational efficiency and model editability. These improvements make the architecture more practical and flexible.

Critical weaknesses

Although the proposed method shows promising results, the current experiments do not demonstrate a clear advantage over existing baselines in terms of editing performance, particularly activation steering methods. This raises questions about the unique benefits of the proposed editing mechanism.

Questions for authors

Given that conventional Transformers can already achieve comparable or even better behavioral editing results through activation steering, what unique advantages does ProtoEdit offer over these existing approaches?

1 comment

We thank the reviewer for their time and suggestions! We will add any new comments and presentation improvements to the camera-ready version of the paper upon acceptance.


Weakness/Question: Although the proposed method shows promising results, the current experiments do not demonstrate a clear advantage over existing baselines in terms of editing performance, particularly activation steering methods. This raises questions about the unique benefits of the proposed editing mechanism. Given that conventional Transformers can already achieve comparable or even better behavioral editing results through activation steering, what unique advantages does ProtoEdit offer over these existing approaches?

We agree that ProtoEdit does not outperform the strongest activation-steering methods on raw edit success, and we do not claim that it does: it is competitive with them (e.g., 0.973 vs. 0.964 for DiffMean on Neg→Pos).

ProtoEdit's distinct contribution is that it is architecture-native: rather than adding a fixed direction to the residual stream, it inserts the edit as an explicit synthetic prototype that competes with the learned prototypes in the model's own write-gate routing, so the edit is applied in proportion to how strongly each token routes to it. This shows that the prototype mechanism is a practical place to edit model behavior, not only to inspect it. The routing competition matters: with the same records, a variant that scores the same gate outside the write-gate softmax ends at twice ProtoEdit's toxicity (0.095 vs. 0.047) and at higher disfluency (14.34 vs. 11.85) (Appendix A.7), and ProtoEdit has the lowest perplexity among all arms above 0.85 on Pos→Neg. Unlike the subtractive prototype editing of prior work, it is also not limited to concepts the model has already isolated in a prototype.

Our main finding from the editability results is that: ProtoT v2 is a more editable backbone than LLaMA-3. To test this, we applied the same steering methods (ActAdd and the four AxBench recipes), with an identical tuning and selection protocol, to ProtoT v2 and to a parameter-matched LLaMA-3 trained on the same data with the same tokenizer (Section 5.2, Table 4). Measured as gain over each backbone's own unedited model (Appendix Table 22), the same method gives a larger gain on ProtoT v2 in 10 of 15 (method, condition) comparisons, a smaller one in 3, with 2 indistinguishable. Using non-overlapping 95% prompt-level bootstrap CIs on these gains as a conservative criterion, ProtoT v2 is significantly better for ActAdd on all three conditions, for PCA on both sentiment directions, and for DiffMean and LAT on detoxification. The gaps can be large: ActAdd flips 0.745 vs. 0.472 of reviews on Pos→Neg, and reaches 0.038 vs. 0.102 toxicity. On detoxification, DiffMean, LAT and ActAdd reach 40–63% lower toxicity on ProtoT v2, even though its unsteered baseline is slightly more toxic (0.2315 vs. 0.2157). The advantage is not uniform (LLaMA-3 is better for DiffMean on Neg→Pos and for Probe and LAT on Pos→Neg), so we read it as method-dependent rather than universal. Since none of these baselines access the prototypes, this suggests that part of ProtoT v2's editability lies in its representations themselves, and conventional steering methods benefit from it directly.

Reviewer_WtC3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper improves the prototype transformer, which replaces self-attention with prototype-based recurrent communication channels. ProtoTv2 uses a stable recurrence, simplifies read routing, removes local causal convolutions, and introduces a custom Triton kernel. It also proposes ProtoEdit, which uses a synthetic prototype to gate concept-direction injections into the value stream. At approximately 80M parameters, the method improves perplexity over the original ProtoT at 1024/2048-token contexts. At approximately 400M parameters, it has higher training throughput but worse perplexity than the comparison architectures. Sentiment editing and detoxification are competitive with strong steering baselines. Prototype interpretability is supported mainly by activation-based semantic evaluation.

Critical strengths

The paper improves an interpretable architecture. Replacing the scaled-cumsum computation with a stable recurrence addresses a specific numerical failure mode. At the base scale, perplexity improves from 28.05 to 25.89 at context 1024, and the optimized implementation gives substantial measured training-speed gains.

The editing evaluation includes strong simple baselines and useful quality controls. ProtoEdit raises sentiment-flip success to 97.3% and 87.4% in the two directions and reduces mean toxicity from 0.2315 to 0.0474.

Critical weaknesses

The evidence does not establish that the readable prototypes faithfully identify the concepts responsible for predictions. Table 5 scores themes in highly activated text. However, Appendix A.8 shows that masking the corresponding prototype does not consistently suppress the concept. It can even increase the token probability. ProtoEdit demonstrates steerability, but its injected concept direction does not by itself validate the causal meaning of the learned prototypes. This limits a central motivation for the architecture.

The practical speed-quality tradeoff is not sufficiently established. At approximately 400M parameters, ProtoT v2 has perplexity 14.47 versus 12.90 for LLaMA-3 and 12.93 for GDN-2, with substantial retrieval deficits. Table 8 also combines stabilization and the kernel in one ablation, limiting attribution of the quality gain.

Questions for authors

Can the authors test whether a prototype's assigned concept predicts the effect of intervening on it? Please use held-out examples and controls for redundant channels or bypass paths.

3 comments

(Part 1 of n)

We thank the reviewer for their time and suggestions! We will add any new results and presentation improvements to the camera-ready version of the paper upon acceptance.


Weakness 1: The evidence does not establish that the readable prototypes faithfully identify the concepts responsible for predictions. Table 5 scores themes in highly activated text. However, Appendix A.8 shows that masking the corresponding prototype does not consistently suppress the concept. It can even increase the token probability. ProtoEdit demonstrates steerability, but its injected concept direction does not by itself validate the causal meaning of the learned prototypes. This limits a central motivation for the architecture.

Our additional experiments (in our response to your Question below) show clear reductions in concept-token probabilities when predictions require using concept information from earlier context. For example, in “The invitation said, ‘The meeting will take place in January.’ The meeting month is January,” masking the original month prototype reduces the final January token’s probability by 26.1%, with the same intervention method as in Appendix A.8. Multi-prototype interventions strengthen these effects, while cross-concept controls show that intervening on different-concept groups does not produce comparable reductions. These results provide causal evidence linking the learned prototypes’ semantic labels to their role in using contextual information. Detailed results and controls are provided in our response to your Question below.

Weakness 2 (a): The practical speed-quality tradeoff is not sufficiently established. At approximately 400M parameters, ProtoT v2 has perplexity 14.47 versus 12.90 for LLaMA-3 and 12.93 for GDN-2, with substantial retrieval deficits.

To evaluate the speed-quality trade-off, we trained all 5 models at the base scale (80-100M parameters) with approximately equal wall-clock (time) budget on the same hardware. The full experiment is presented in our answer to the Question from Reviewer 8Z2x. The results show ~10 percentage point less gap versus Mamba-3 and GDN-2 at the base scale, which we estimate would translate to similar performance to Mamba-3 and GDN-2 at the large scale. This would mean that ProtoT v2 would perform similarly to the linear-compute baselines under the same budget at the large scale. As for LLaMA, while it is the fastest baseline at this scale, its quadratic compute self-attention leads to relatively slower speeds at larger context lengths, as discussed in Appendix A.5 Throughput at Large Contexts.

Beyond the speed–quality trade-off, we would also like to highlight a separate contribution that may have been overlooked: ProtoT v2 demonstrates stronger editability as a backbone than LLaMA-3. When applying the same five conventional steering methods to both models under an identical evaluation protocol, the same method gives a larger gain on ProtoT v2 than on LLaMA-3 in 10 of 15 evaluated settings (3 smaller, 2 indistinguishable) (Section 5.2, Table 4; Appendix A.7, Table 22). Importantly, none of these methods accesses the prototype bank, suggesting that ProtoT v2's representations themselves are particularly amenable to steering, independently of our “ProtoEdit” method.

Weakness 2 (b): Table 8 also combines stabilization and the kernel in one ablation, limiting attribution of the quality gain.

To address your concern, we implemented the stabilization (recurrent formulation) for the ProtoT baseline via a simple matmul chunked scan (which computes it exactly) to make the recurrent formulation’s speed somewhat manageable (3.9 it/s vs 5.0 it/s for the ProtoT baseline), and ran it with 3 LRs (on the same LR grid). The results (below) show that the recurrent time-discount formulation obtains approximately the same performance as the ProtoT v2 kernel (when all other settings of Table 8 are set to same as ProtoT), which confirms our claim that the performance gains of the ProtoT v2 kernel are entirely due to the numeric stabilization:

ModelFinal PPL
ProtoT28.05
new: ProtoT (recurrent time-discount)25.74
ProtoT v2 (with kernel, but otherwise same settings as ProtoT)25.79
ProtoT v225.89

(Part 2 of 3)

Question: Can the authors test whether a prototype's assigned concept predicts the effect of intervening on it? Please use held-out examples and controls for redundant channels or bypass paths.

Yes. We ran additional experiments on held-out examples where predictions depend on concept information provided earlier in the context. Prototype labels and groups were selected using activations on separate training texts, without using the evaluation examples.

Single-prototype interventions. Table 1 shows results using the original write-gate masking, read-gate masking, and re-initialization operations. The tested prototypes are the same color, month, and nationality prototypes reported in Appendix A.8. As shown in the table, with the same prototypes and WriteMask operator, the probabilities of January and March decrease by 26.1% and 19.6%, respectively. These results show that the learned prototypes contribute to predictions involving their assigned concepts when the relevant information is provided in context.

Table 1. Subtractive editing on context-dependent examples in ProtoT v2:

HubTypeSentenceBaseΔp W / R / I (%)Δppl (%)
L0P23 (colors)gen.The door was painted red. The color of the door was red.10.10%−3.8 / −6.0 / −10.0−5.4
L0P23 (colors)gen.The door was painted blue. The color of the door was blue.5.68%+3.8 / +3.3 / −6.9−5.1
L0P52 (months)gen.The invitation said, “The meeting will take place in January.” The meeting month is January.40.53%−26.1 / −2.5 / −1.2−3.0
L0P52 (months)gen.The invitation said, “The meeting will take place in March.” The meeting month is March.40.36%−19.6 / −0.5 / −2.5−3.9
L0P22 (countries)gen.The visitor said, “I am from France.” The visitor’s home country is France.22.66%−2.0 / +0.6 / +1.6−0.5
L0P22 (countries)gen.The visitor said, “I am from Germany.” The visitor’s home country is Germany.22.71%−3.4 / +0.1 / +3.4−0.3

Controls for redundant channels or bypass paths. Other prototypes may carry related information, allowing the model to retain access to a concept after a single channel is masked. To test this possibility, we applied the same interventions jointly to three prototypes associated with each concept (Table 2).

Table 2. Multi-prototype joint Subtractive Editing of Prototype groups on ProtoT v2:

Prototype groupTypeSentenceBaseΔp W / R / I (%)Δppl (%)
Colors: L0P23, L7P47, L17P11gen.The door was painted red. The color of the door was red.10.10%−19.8 / −21.7 / −20.2−4.4
Colors: L0P23, L7P47, L17P11gen.The door was painted blue. The color of the door was blue.5.68%−6.7 / −7.3 / −11.5−4.8
Months: L0P52, L15P33, L7P37gen.The invitation said, “The meeting will take place in January.” The meeting month is January.40.53%−59.3 / −48.1 / −63.6+0.7
Months: L0P52, L15P33, L7P37gen.The invitation said, “The meeting will take place in March.” The meeting month is March.40.36%−45.5 / −34.2 / −57.2−1.5
Countries: L1P42, L0P22, L15P20gen.The visitor said, “I am from France.” The visitor’s home country is France.22.66%−13.5 / −11.4 / −12.9−0.0
Countries: L1P42, L0P22, L15P20gen.The visitor said, “I am from Germany.” The visitor’s home country is Germany.22.71%−11.6 / −6.4 / −4.9+0.4

As shown in Table 2, these group interventions produced larger reductions in target-token probability than the single-prototype interventions in Table 1. For example, jointly masking the month group at the write gate reduced the probabilities of January and March by 59.3% and 45.5%, compared with 26.1% and 19.6% when masking L0P52 alone.

(Part 3 of 3)

Table 3 provides further cross-concept validation for the presented examples. Intervening on different concept groups did not result in an equivalent reduction in target-token probability: masking the "color" group had almost no impact on "month" predictions, while masking the "month" group had only a very limited effect on "country" predictions. These control analyses support concept-selective effects of the prototype groups rather than simply stemming from a general loss of predictive capacity caused by the intervention on the three channels.

Table 3. Cross-Concept Controls for Multi-prototype joint Subtractive Editing on ProtoT v2:

Sentence categoryIntervened prototype groupSentenceBaseΔp W / R / I (%)Δppl (%)
ColorsCountries: L1P42, L0P22, L15P20The door was painted red. The color of the door was red.10.10%+7.88 / +8.02 / +4.58−1.21
ColorsCountries: L1P42, L0P22, L15P20The door was painted blue. The color of the door was blue.5.68%+3.89 / +4.06 / +2.99−0.85
MonthsColors: L0P23, L7P47, L17P11The invitation said, “The meeting will take place in January.” The meeting month is January.40.53%−0.02 / −0.05 / +0.01−0.03
MonthsColors: L0P23, L7P47, L17P11The invitation said, “The meeting will take place in March.” The meeting month is March.40.36%+0.00 / −0.03 / +0.02−0.05
CountriesMonths: L0P52, L15P33, L7P37The visitor said, “I am from France.” The visitor’s home country is France.22.66%−0.29 / +0.04 / −0.14+0.03
CountriesMonths: L0P52, L15P33, L7P37The visitor said, “I am from Germany.” The visitor’s home country is Germany.22.71%−0.20 / +0.01 / −0.10+0.03

Together, these held-out experiments show that the assigned concepts predict which token probabilities are more strongly affected by intervention. Multi-prototype interventions reveal stronger effects across related channels, while cross-concept controls support a concept-specific contribution to the model’s use of contextual information.

Reviewer_61gCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ProtoT v2, an improved version of ProtoT with a numerically stable recurrent formulation, several architectural simplifications, and a custom Triton kernel. These changes substantially improve training and inference efficiency and reduce memory usage, while enabling evaluation at much larger model and data scales. The paper also introduces ProtoEdit, a synthetic-prototype method for targeted model editing, and evaluates interpretability and robustness.

Critical strengths
  1. The paper provides a substantial practical improvement over the original ProtoT implementation. The numerically stable recurrent formulation and custom Triton kernel substantially improve speed, memory efficiency, and scalability. The larger-scale evaluation against recent linear-compute baselines also provides a more realistic assessment of ProtoT than the original small-scale experiments.

  2. ProtoEdit provides a useful prototype-level editing interface. It extends the original ProtoT editing mechanism beyond concepts already represented by existing prototypes, enabling targeted edits through synthetic prototypes. The results on sentiment editing and detoxification are competitive with strong activation-steering baselines and suggest that ProtoT v2 is a relatively editable backbone.

Critical weaknesses
  1. The main weakness is the limited conceptual advance over ProtoT. ProtoT v2 retains the core prototype-mediated communication mechanism of ProtoT, while most changes simplify, stabilize, or optimize the existing formulation; the ablations further show that most of the perplexity gain comes from numerical stabilization, with the kernel providing the main speed and memory gains. These are useful practical improvements, but it remains unclear what new modeling principle ProtoT v2 contributes beyond making ProtoT more efficient and scalable.

  2. The evidence for the claimed native interpretability remains limited. The evaluation is largely based on top-activating examples scored by an LLM, and ProtoT v2 still underperforms the SAE baseline on all three interpretability metrics. More importantly, routing all cross-token communication through prototypes does not by itself show that these channels are semantically stable, coherent, or causally faithful.

Questions for authors
  1. What is the main new scientific or modeling contribution of ProtoT v2 beyond ProtoT? If ProtoT already used the numerically stable recurrence and an optimized kernel, what substantive contribution would remain beyond ProtoEdit?
  2. Can the authors provide stronger evidence that the architectural simplifications preserve or improve interpretability relative to ProtoT? A matched ProtoT/ProtoT v2 comparison, especially with and without the causal convolution, would be helpful.
  3. How should the scalability claim over ProtoT be interpreted given that ProtoT is not evaluated at the large scale? Please clarify which scalability gains are directly demonstrated and which are inferred from the implementation.
4 comments

(Part 1 of n)

We thank the reviewer for their time and suggestions! We will add any new results and presentation improvements to the camera-ready version of the paper upon acceptance.


Weakness 1: The main weakness is the limited conceptual advance over ProtoT. ProtoT v2 retains the core prototype-mediated communication mechanism of ProtoT, while most changes simplify, stabilize, or optimize the existing formulation; the ablations further show that most of the perplexity gain comes from numerical stabilization, with the kernel providing the main speed and memory gains. These are useful practical improvements, but it remains unclear what new modeling principle ProtoT v2 contributes beyond making ProtoT more efficient and scalable.

ProtoT v2 removes several components that are part of the ProtoT architecture, and reformulates its time-discounted aggregation into a numerically stable recurrent formulation. As mentioned throughout the paper, one notable conceptual advance is that removing the causal convolution makes the prototype mechanism the only mechanism for exchanging information between sequence positions. This makes prototypes have a central role in the positional mixing. Furthermore, in our answer to your Question 2, we show that ProtoT v2 has similar or slightly better prototype interpretability than ProtoT.

The main contributions also include:

  1. ProtoEdit, a new synthetic-prototype editing mechanism (Section 5.2, Table 4). Beyond achieving competitive editing performance, ProtoEdit demonstrates that the prototype mixer provides a natural location for concept-level steering, allowing interventions to operate directly through the model’s existing token-mixing mechanism.

  2. ProtoT v2 has better editability as a backbone than LLaMA-3: across five standard steering methods applied to both models under the same evaluation protocol, it gives a larger gain over the unedited model in 10 of 15 (method, condition) comparisons, a smaller one in 3, with 2 indistinguishable (Section 5.2, Table 4; and Appendix A.7, Table 22).

  3. Finally, our updated existing-prototype editing results (answer to Question, reviewer WtC3) provide evidence that the prototype semantic interpretations are causally meaningful: intervening on groups of concept-related prototypes selectively affects predictions associated with those concepts.

(Part 2 of n)

Weakness 2: The evidence for the claimed native interpretability remains limited. The evaluation is largely based on top-activating examples scored by an LLM, and ProtoT v2 still underperforms the SAE baseline on all three interpretability metrics. More importantly, routing all cross-token communication through prototypes does not by itself show that these channels are semantically stable, coherent, or causally faithful.

Evaluation protocol. Scoring top-activating examples with an LLM is the standard automated-interpretability protocol used to evaluate SAE features (Bricken et al., 2023; Paulo et al., 2025). We apply it identically to every method in Table 5, including the SAE baseline, so the comparison is like-for-like. Within this protocol, ProtoT v2's prototypes score clearly above a null baseline that assigns random sentences to each prototype (disentanglement 2.73 vs. 2.02, coverage 2.56 vs. 1.32, number of themes 7.42 vs. 9.01). The matched base-scale comparison in our answer to your Question 2 confirms this: ProtoT, ProtoT v2, and ProtoT v2 with causal convolution are all significantly above the null (p < 10⁻²⁷).

Comparison with the SAE. The SAE is a post-hoc method: it is trained separately on top of LLaMA-3's residual stream and only approximates the model's computation, with a reconstruction error. ProtoT v2's prototypes are instead part of the model's computation: they are the channels through which all cross-token communication actually flows, with no additional training. The like-for-like comparison for native interpretability is therefore with LLaMA-3's own components, its attention heads, and ProtoT v2 outperforms them on all three metrics (2.73 vs. 2.38 disentanglement, 2.56 vs. 2.08 coverage, 7.42 vs. 8.08 themes). That the prototypes come close to, but stay below, a dedicated post-hoc method is consistent with our claim. We will make this distinction explicit in the paper.

Causality / Faithfulness. Routing all cross-token communication through prototypes makes them structurally faithful: the routing weights we inspect are exactly the ones the model uses to move information between positions, which is why we removed the causal convolution. However, this does not establish that each prototype's assigned theme is causally faithful. As Appendix A.8 shows, ablating a single prototype has inconsistent effects. To address this concern, we ran additional experiments (Response to reviewer WtC3’s Question) that show clear reductions in concept-token probabilities when predictions require using concept information from earlier context. For example, masking the month prototype reduces the probabilities of January and March by 26.1% and 19.6%. Jointly intervening on three same-concept prototypes strengthens these effects and reduces target-token probabilities across all examples under all three intervention operators. Intervening on different-concept groups does not produce comparable reductions. These results provide causal evidence that the tested prototypes contribute to the model’s use of concept information from context. These findings link the prototypes’ semantic labels to their causal role in using contextual information, extending the evidence beyond LLM-rated activation coherence.

(Part 3 of 3)

Question 1: What is the main new scientific or modeling contribution of ProtoT v2 beyond ProtoT? If ProtoT already used the numerically stable recurrence and an optimized kernel, what substantive contribution would remain beyond ProtoEdit?

We address this question in more detail in Weakness 1 above. Apart from the numerically stable recurrence, optimized kernel, and ProtoEdit, ProtoT v2 has a conceptually simpler architecture, by removing a number of architectural choices that have little to no impact on performance. Most notably, we remove the local causal convolution, thus leaving the prototype mechanism as the only mechanism for exchanging information between sequence positions. This is aligned with the goal of making the prototypes the primary component of the mixer to inspect, interpret, and intervene on.

In our answer to Weakness 1 above we also list a number of our editability-related contributions beyond ProtoEdit.

Question 2: Can the authors provide stronger evidence that the architectural simplifications preserve or improve interpretability relative to ProtoT? A matched ProtoT/ProtoT v2 comparison, especially with and without the causal convolution, would be helpful.

As pointed out in the paper, we could not train the ProtoT baseline at the large scale (which is used for this experiment) due to computational constraints. To accommodate your request, we thus decided to do this comparison on the base scale: ProtoT, ProtoT v2, and ProtoT v2 (with causal conv).

All three models have 12 layers, width 512 and 64 prototypes per layer (768 in total), and were trained on the same ClimbMix data with the same tokenizer and seed; ProtoT v2 (with causal conv) differs from ProtoT v2 only by the causal convolution in the first two layers. We use the same interpretability pipeline as in Table 5 of the paper, with the same LLM judge for every row (DeepSeek-V4), so absolute values are not directly comparable to Table 5.

ModelDisentanglement ↑Coverage (Main Theme) ↑# Themes ↓
ProtoT2.922 ± 1.5302.702 ± 2.1627.266 ± 2.432
ProtoT v2 (with causal conv)2.888 ± 1.3712.720 ± 2.0397.268 ± 2.366
ProtoT v22.954 ± 1.4302.852 ± 2.2097.099 ± 2.415
ProtoT v2 null2.220 ± 0.4971.576 ± 0.7078.603 ± 1.418

Interpretation of reported variability: the standard deviations in this table measure variability across individual prototypes (or corresponding baseline features), rather than across experimental runs. Thus, they reflect differences in interpretability between individual features, not uncertainty in the mean scores. The means provide a direct comparison of average interpretability across models, although variability across training seeds or repeated LLM evaluations is not assessed.

The differences between the three models are not statistically significant on any metric, while all three are significantly above the null baseline. The architectural simplifications therefore preserve the interpretability of ProtoT's prototypes, and removing the causal convolution does not reduce it; ProtoT v2 is nominally the best on all three metrics, although the improvement is not significant.

Question 3: How should the scalability claim over ProtoT be interpreted given that ProtoT is not evaluated at the large scale? Please clarify which scalability gains are directly demonstrated and which are inferred from the implementation.

For scalability, we show improvements in two axes:

  1. Training with longer context (context scalability), which we demonstrated in Section 5.1 Performance and Scalability: ProtoT v2 pulls ahead of ProtoT at context lengths above 512: 1024 and 2048, likely due to its numerically-stable time-discount formulation.

  2. Speed/compute efficiency: this is a practical advantage, where for a given time/compute budget, we can train substantially larger models for ProtoT v2, because it has approximately one-third of the memory usage of ProtoT, and about three times faster training and inference speed.

In our answer to the Question from reviewer 8Z2x, we also show that at the base scale, the speed of ProtoT v2 allows for about three times more training tokens to be used versus ProtoT, resulting in 25% lower perplexity (24.61 vs 32.80) at equal wall-clock time.

Thank you for the detailed response. However, my main concern about novelty remains: I view the removal of the causal convolution and the other architectural changes primarily as simplifications and practical improvements rather than a substantial new modeling principle beyond ProtoT. So I maintain my score.

Reviewer_8Z2xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents ProtoT v2, a prototype-based language model with a simplified mixer, a numerically stable recurrence, and a custom Triton implementation. It also introduces ProtoEdit, which uses synthetic prototypes to steer generation without retraining the model.

Critical strengths
  1. The numerical stabilization and kernel optimization deliver substantial practical improvements over ProtoT. These changes make the architecture more viable at larger scales.

  2. ProtoEdit provides effective sentiment steering and detoxification. I appreciate the comparisons against simple editing baselines and the use of held-out tuning with fluency, diversity, and relevance constraints.

Critical weaknesses
  1. The interpretability evidence is weaker. Table 5 measures LLM-rated semantic consistency of highly activating text, which does not establish that the assigned concepts faithfully explain the model's computation. Removing semantically coherent prototypes produces inconsistent or negligible effects on the corresponding predictions. ProtoEdit's success shows controllability, but does not resolve this gap in understanding the learned prototypes.

  2. I am not yet convinced that the benefits justify the loss in capability. At 400M parameters, perplexity higher than LLaMA-3, and several retrieval tasks show much larger gaps. ProtoEdit is also not consistently better than simple DiffMean editing.

Questions for authors

How does ProtoT v2 compare with the baselines when all models receive the same wall-clock training budget on the same hardware?

2 comments

(Part 1 of 2)

We thank the reviewer for their time and suggestions! We will add any new results and presentation improvements to the camera-ready version of the paper upon acceptance.


Weakness 1: The interpretability evidence is weaker. Table 5 measures LLM-rated semantic consistency of highly activating text, which does not establish that the assigned concepts faithfully explain the model's computation. Removing semantically coherent prototypes produces inconsistent or negligible effects on the corresponding predictions. ProtoEdit's success shows controllability, but does not resolve this gap in understanding the learned prototypes.

We agree that Table 5 (in the paper) measures semantic coherence of highly activating text and does not, by itself, establish causal faithfulness. To address this gap, we now extend our evaluation of learned prototypes while still using the subtractive interventions described in Appendix A.8 (see our response to reviewer WtC3’s Question for detailed results and examples). As a reminder, Appendix A.8 of the paper reported mixed effects for the color prototype and negligible effects for the month prototype on the tested sentence completions.

We now evaluate these same prototypes on paired examples where the answer must track information supplied earlier in the context. For example, in “The invitation said, ‘The meeting will take place in January.’ The meeting month is January,” masking the original month prototype L0P52 reduces the final January token’s probability by 26.1%. Multi-prototype interventions (jointly intervening on three month-related prototypes) strengthen these effects to 59.3%, while cross-concept controls show that intervening on different-concept groups does not produce comparable reductions.

These experiments further link the semantic consistency observed in Table 5 (of the paper) to the model’s actual computations: the conceptual labels assigned to prototypes can predict which channels, when intervened upon, will substantially affect the predicted probability of the corresponding concept words. Jointly intervening on multiple prototypes associated with the same concept simultaneously yields a stronger effect than intervening on a single prototype; meanwhile, the control experiments show little or no reduction in target-token probabilities when intervening on prototypes associated with other concepts, supporting that the observed reductions are tied to the corresponding concepts rather than general disruption of the model. Consequently, in these contextual tasks, semantic labels not only characterize which texts activate a given prototype but also reflect the functional role that the corresponding group of prototypes plays as the model processes contextual information.

Weakness 2 (a): I am not yet convinced that the benefits justify the loss in capability. At 400M parameters, perplexity higher than LLaMA-3, and several retrieval tasks show much larger gaps.

We address the “loss in capability” in our answer to your Question below. Beyond training speed and linear computational scalability (versus LLaMA-3’s quadratic), the benefits of ProtoT v2 include:

  1. ProtoT v2 offers a more directly inspectable token-mixing mechanism through its learned prototypes, which score better than LLaMA-3’s attention heads on our interpretability metrics (Section 5.3, Table 5). Moreover, our updated existing-prototype editing results (above, and in answer to the Question of reviewer WtC3) provide evidence that these semantic interpretations are causally meaningful: intervening on groups of concept-related prototypes selectively affects predictions associated with those concepts.

  2. ProtoT v2 also provides an architecture-native interface for concept steering: ProtoEdit introduces synthetic prototypes into the mixer, enabling targeted interventions through the model's own routing mechanism (Section 5.2, Table 4).

  3. ProtoT v2 demonstrates stronger editability as a backbone than LLaMA-3: across five standard steering methods applied to both models under the same evaluation protocol, it gives a larger gain over the unedited model in 10 of 15 (method, condition) comparisons, a smaller one in 3, with 2 indistinguishable (Section 5.2, Table 4; and Appendix A.7, Table 22).

(Part 2 of 2)

Weakness 2 (b): ProtoEdit is also not consistently better than simple DiffMean editing.

We agree, and we do not claim that ProtoEdit outperforms DiffMean – the two methods are complementary rather than ranked. They are on par on Neg→Pos (0.973 vs. 0.964 at similar perplexity, 9.62 vs. 9.44). On Pos→Neg and detoxification, DiffMean edits more strongly (0.916 vs. 0.874; 0.0245 vs. 0.0474 toxicity), while ProtoEdit produces more fluent text (perplexity 11.24 vs. 14.48 and 11.85 vs. 12.34).

Our main editability claim concerns the backbone rather than a single method: the same steering methods, tuned with an identical protocol, give the better edit on ProtoT v2 than on a parameter-matched LLaMA-3 in 10 of 15 comparisons (a smaller one in 3, with 2 indistinguishable) (Section 5.2, Table 4; and Appendix A.7, Table 22). This includes DiffMean itself (0.916 vs. 0.853 on Pos→Neg; 0.0245 vs. 0.0542 toxicity).

ProtoEdit's contribution is to make this editability available through a prototype-level interface: the edit is an explicit prototype that competes in the model's own routing, and is not restricted to concepts the model has already isolated in a prototype. This shows that the prototype mechanism is a practical place to edit model behavior, not only to inspect it.

Question: How does ProtoT v2 compare with the baselines when all models receive the same wall-clock training budget on the same hardware?

Due to computation costs, we do this comparison at the base (80-100M parameter) scale. We record the original base-scale experiment wall-clock times (which range from 1h to 3h) and derive the number of training tokens for each model to achieve approximately equal wall-clock time of 1.5h on 1 H100 GPU. We train each model on 3 LRs (on the same LR grid as in the paper), and check that the best perplexity value is achieved at the middle LR value.

ModelFinal dev PPLWall-clock (min)it/s
LLaMA-322.0891.811.60
Mamba-322.0391.38.42
GDN-222.6592.96.78
ProtoT32.8085.05.03
ProtoT v224.6190.714.51

ProtoT v2 at matched wall-clock time is: 11.4%, 11.7%, and 8.6% worse perplexity than LLaMA-3, Mamba-3, and GDN-2, respectively. Comparing these gaps to the data-matched base scale gaps from the paper (Section 5.1 Table 1): 13.9%, 20.5%, and 22.24% worse perplexity than LLaMA-3, Mamba-3, and GDN-2, respectively, suggests improvements in perplexity versus Mamba-3 and GDN-2 by about 10% from the data-matched to the new wall-clock-matched setup.

Extrapolating these gains to the large scale (where in the data-matched experiment, Section 5.1 Table 2, ProtoT v2 has 12.2%, 6.2%, and 11.9% worse perplexity than LLaMA-3, Mamba-3, and GDN-2, respectively), we would expect ProtoT v2 to perform roughly on par with Mamba-3 and GDN-2 in a wall-clock-matched setup.

In conclusion, we expect the wall-clock-adjusted perplexity numbers at large scale to fall roughly on par with Mamba-3 and GDN-2 (which are also linear-compute models), but still ~10% behind LLaMA (which uses quadratic-compute self-attention). Note: while LLaMA is the fastest baseline model at ctx=1024, it falls behind speed-wise at large context sizes, due to the self-attention dominating the computation (see Appendix A.5 Throughput at Large Contexts, Table 15).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.