pith. machine review for the scientific record. sign in

arxiv: 2604.12426 · v1 · submitted 2026-04-14 · 💻 cs.LG · cs.CL

Recognition: no theorem link

Do Transformers Use their Depth Adaptively? Evidence from a Relational Reasoning Task

Alicia Curth , Rachel Lawrence , Sushrut Karmalkar , Niranjani Prasad
Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:30 UTC · model grok-4.3

classification 💻 cs.LG cs.CL
keywords transformersadaptive depthrelational reasoninglogit lenscausal patchingfinetuningmulti-hop reasoning
0
0 comments X

The pith

Finetuned transformers show clearer adaptive depth use on relational reasoning tasks than pretrained ones.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tests whether transformers allocate their layers differently depending on how many relationship hops a reasoning task requires. They track this by reading out predictions layer by layer with the logit lens and by measuring when and how information from different tokens gets combined using causal patching. Pretrained models give only weak signals of adaptation: some larger models reach reasonable answers earlier on short chains, and most models pull in information from more layers as chains grow longer. After finetuning on the same task the pattern becomes stronger and more reliable, especially when the finetuning is allowed to change the model away from its original language-modeling behavior.

Core claim

On a controlled family-relation reasoning task whose difficulty scales with the number of hops that must be composed, logit-lens readouts and causal-patching measurements reveal limited adaptive depth in pretrained transformers, with larger models sometimes converging on easier instances in fewer layers and all models generally recruiting more layers to integrate information as hop count rises. The same measurements on task-finetuned models yield clearer and more consistent evidence that depth usage scales with difficulty, and the effect is larger when finetuning is less constrained and therefore moves the model farther from its original language-modeling distribution.

What carries the argument

Logit-lens early readouts that expose how the final prediction distribution evolves across successive layers, together with causal-patching interventions that quantify the number of layers required to integrate tokens that hold successive links in a relational chain.

Load-bearing premise

The logit lens and causal patching measurements accurately reflect whether the model is genuinely using different amounts of depth rather than merely producing different surface-level output patterns.

What would settle it

If, after finetuning, the logit lens shows essentially identical prediction trajectories for short and long chains and causal patching shows no systematic increase in the layers needed to combine tokens as hop count grows, the claim of adaptive depth use would be falsified.

Figures

Figures reproduced from arXiv: 2604.12426 by Alicia Curth, Niranjani Prasad, Rachel Lawrence, Sushrut Karmalkar.

Figure 1
Figure 1. Figure 1: Example CLUTRR stories for 2-hop and 5-hop reasoning. Each sentence states a family relation between two people. The model must compose the chain of relations to infer the target relationship between the query pair. show that layers in the second half of the model contribute substantially less to the residual stream than those in the first half and that skipping later layers has a smaller effect on predict… view at source ↗
Figure 2
Figure 2. Figure 2: Probability assigned to family relation tokens when decoding the hidden states of the final token using the language modeling head directly, by layer across different pretrained model sizes and families, averaged across hops. Family relations are decodable at layer l < L across all models. spanning a range of sizes up to 14B parameters. The upper end of this range is determined by the computational cost an… view at source ↗
Figure 3
Figure 3. Figure 3: Logit lens results for Phi and Qwen2.5 families. Correctness, constrained correctness and probability assigned to family relation tokens by logit lens predictions by layer, colored by hops. X-axes (layers) are left-truncated for better readability; previous accuracies & p fam l are zero. (a) Average recovery score by replaced relation across different hops (rows), colored by relation replaced. (b) Average … view at source ↗
Figure 4
Figure 4. Figure 4: Causal patching results. Average recovery score at the replaced token t r (dashed lines) and the final token T (solid lines) by model depth. longer stories (especially relationship tokens in the middle of longer stories in panel (a)) as indicated by recovery scores dropping at earlier layers, meaning that information is mixed into the residual streams of future tokens sooner for longer stories6 . Interesti… view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy of the top logit lens prediction across models (columns) and training regimes (rows) by layer and colored by number of hops. last. Qwen2.5 models appear to fully integrate information about the first relationship at T later than Phi models for longer stories. Overall, the results in this section indicate that pretrained models do show some variation in depth use across task difficulty: larger mode… view at source ↗
Figure 6
Figure 6. Figure 6: Full causal patching and logit-lens trajectories for a 5-hop example of finetuned GPT2-large versions. (making it possible in principle to solve the original task by memorising relations by name rather than reasoning about them) we rename all individuals to Person1 through PersonK and randomise the order in which numerical indices appear. Further implementation details are provided in Ap￾pendix A. • Observ… view at source ↗
Figure 7
Figure 7. Figure 7: Causal patching results for GPT2-large, pretrained and finetuned. Average recovery score at the replaced token t r (dashed lines) and the final token T (solid lines) by model depth. difficulty. The fully finetuned models tell a different story: accuracy for harder tasks rises later in the network, especially for larger models, suggesting that these models do learn to process the task more iteratively with … view at source ↗
read the original abstract

We investigate whether transformers use their depth adaptively across tasks of increasing difficulty. Using a controlled multi-hop relational reasoning task based on family stories, where difficulty is determined by the number of relationship hops that must be composed, we monitor (i) how predictions evolve across layers via early readouts (the logit lens) and (ii) how task-relevant information is integrated across tokens via causal patching. For pretrained models, we find some limited evidence for adaptive depth use: some larger models need fewer layers to arrive at plausible answers for easier tasks, and models generally use more layers to integrate information across tokens as chain length increases. For models finetuned on the task, we find clearer and more consistent evidence of adaptive depth use, with the effect being stronger for less constrained finetuning regimes that do not preserve general language modeling abilities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper investigates whether transformer models use their depth adaptively on tasks of varying difficulty, using a controlled multi-hop relational reasoning task based on family relationship chains. Difficulty is parameterized by the number of hops. The authors track prediction evolution via logit lens early readouts and measure cross-token information integration via causal patching experiments. They report limited evidence of adaptive depth use in pretrained models (e.g., larger models stabilizing earlier on short chains) and clearer, more consistent evidence in finetuned models, with stronger effects under less constrained finetuning that does not preserve general language modeling capabilities.

Significance. If the central observations hold under tighter controls, the work would provide useful empirical evidence that finetuning regimes can promote adaptive computation in transformers, with potential implications for mechanistic understanding of depth utilization and efficient inference. The controlled task and dual measurement approach (logit lens plus patching) are strengths that allow systematic variation of difficulty; however, the observational character limits the strength of causal claims about internal adaptive mechanisms.

major comments (2)
  1. [§4] §4 (Logit Lens Results): The finding that predictions for shorter chains stabilize at earlier layers does not distinguish adaptive depth allocation from the simpler alternative that easier inputs produce representations that become linearly separable after fewer layers by construction. No controls (e.g., early-exit performance matching full-depth performance on hard examples, or comparison to a fixed-depth baseline) are reported to separate these possibilities.
  2. [§5] §5 (Causal Patching Experiments): The reported increase in cross-token information flow for longer chains could arise from input complexity or token count rather than any internal decision to allocate additional depth. The manuscript lacks ablations such as length-matched controls, shuffled difficulty labels, or forced fixed-layer models that would test whether the patching effect reflects adaptive modulation.
minor comments (2)
  1. [§3] The task description would benefit from a concrete example of a 1-hop vs. 3-hop query in the methods section to make the difficulty parameterization immediately clear to readers.
  2. Figure captions for the patching heatmaps could explicitly note the statistical test and number of examples used to compute the reported effects.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback. The suggestions for tighter controls are well-taken and will help clarify the interpretation of our results on adaptive depth use. We respond to each major comment below and outline the revisions we will incorporate.

read point-by-point responses

  1. Referee: §4 (Logit Lens Results): The finding that predictions for shorter chains stabilize at earlier layers does not distinguish adaptive depth allocation from the simpler alternative that easier inputs produce representations that become linearly separable after fewer layers by construction. No controls (e.g., early-exit performance matching full-depth performance on hard examples, or comparison to a fixed-depth baseline) are reported to separate these possibilities.

    Authors: We agree that the logit lens stabilization pattern for shorter chains is consistent with the alternative that easier inputs simply become linearly separable earlier by construction, without requiring an internal adaptive allocation of depth. Our manuscript already notes that evidence is limited in pretrained models and stronger after finetuning, which we interpret as task-specific encouragement of this behavior. To address the concern directly, we will revise §4 to explicitly discuss this alternative explanation. We will add post-hoc comparisons of early-readout accuracy on hard examples against full-depth performance and include truncated fixed-depth baselines by evaluating the model at intermediate layers. These analyses can be performed without retraining and will be reported in the revision. revision: partial

  2. Referee: §5 (Causal Patching Experiments): The reported increase in cross-token information flow for longer chains could arise from input complexity or token count rather than any internal decision to allocate additional depth. The manuscript lacks ablations such as length-matched controls, shuffled difficulty labels, or forced fixed-layer models that would test whether the patching effect reflects adaptive modulation.

    Authors: We acknowledge that longer chains involve more tokens and greater input complexity, which could drive the observed increase in cross-token patching effects independently of any adaptive depth mechanism. While our task design holds the overall story template fixed and varies only the number of relational hops, token count does scale with difficulty. We will add length-matched controls by subsampling or constructing examples with comparable token lengths across hop counts. We will also include analyses with shuffled difficulty labels to test specificity to actual task difficulty and compare information flow in models forced to fixed shallower depths. These ablations will be incorporated into §5 of the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No significant circularity: purely empirical measurements independent of claims

full rationale

This is an observational study that applies logit-lens readouts and causal patching to track layer-wise prediction stabilization and cross-token information flow on a controlled relational-reasoning task. No derivations, parameter fits, or self-citations are used to generate the central claims; the reported patterns are direct empirical observations whose validity can be assessed against external benchmarks or alternative models. The paper therefore contains no load-bearing steps that reduce to their own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Empirical interpretability study; no free parameters, axioms, or invented entities are introduced. Relies on standard assumptions from prior logit lens and causal intervention literature.

pith-pipeline@v0.9.0 · 5446 in / 1077 out tokens · 41269 ms · 2026-05-10T15:30:14.744923+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 20 canonical work pages · 3 internal anchors

  1. [1]

    Eliciting Latent Predictions from Transformers with the Tuned Lens

    Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens.arXiv preprint arXiv:2303.08112,

    work page internal anchor Pith review arXiv

  2. [2]

    Yangyi Chen, Binxuan Huang, Yifan Gao, Zhengyang Wang, Jingfeng Yang, and Heng Ji

    Dan Biderman, Jacob Portes, Jose Javier Gonzalez Ortiz, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, et al. Lora learns less and forgets less.arXiv preprint arXiv:2405.09673,

    work page arXiv

  3. [3]

    Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901,

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901,

    1901

  4. [4]

    Csordás, C

    R´obert Csord´as, Christopher D Manning, and Christopher Potts. Do language models use their depth efficiently?arXiv preprint arXiv:2505.13898,

    work page arXiv

  5. [5]

    Universal Transformers

    Mostafa Dehghani, Stephan Gouws, Oriol Vinyals, Jakob Uszkoreit, and Łukasz Kaiser. Universal transformers.arXiv preprint arXiv:1807.03819,

    work page internal anchor Pith review arXiv

  6. [6]

    arXiv preprint arXiv:2409.15647 (2024)

    Ying Fan, Yilun Du, Kannan Ramchandran, and Kangwook Lee. Looped transformers for length generalization.arXiv preprint arXiv:2409.15647,

    work page arXiv

  7. [7]

    Nnsight and ndif: Democratizing access to foundation model internals

    Jaden Fiotto-Kaufman, Alexander R Loftus, Eric Todd, Jannik Brinkmann, Caden Juang, Koyena Pal, Can Rager, Aaron Mueller, Samuel Marks, Arnab Sen Sharma, et al. NNsight and NDIF: Democratizing access to foundation model internals.arXiv preprint arXiv:2407.14561,

    work page arXiv

  8. [8]

    Predictability and surprise in large generative models

    10 ICLR 2026 Workshop on Logical Reasoning of Large Language Models Deep Ganguli, Danny Hernandez, Liane Lovitt, Amanda Askell, Yuntao Bai, Anna Chen, Tom Conerly, Nova Dassarma, Dawn Drain, Nelson Elhage, et al. Predictability and surprise in large generative models. InProceedings of the 2022 ACM conference on fairness, accountability, and transparency, ...

    2026

  9. [9]

    Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space

    Mor Geva, Avi Caciularu, Kevin Wang, and Yoav Goldberg. Transformer feed-forward layers build predictions by promoting concepts in the vocabulary space. InProceedings of the 2022 conference on empirical methods in natural language processing, pp. 30–45,

    2022

  10. [10]

    The unreasonable ineffectiveness of the deeper layers.arXiv preprint arXiv:2403.17887, 2024

    Andrey Gromov, Kushal Tirumala, Hassan Shapourian, Paolo Glorioso, and Daniel A Roberts. The unreasonable ineffectiveness of the deeper layers.arXiv preprint arXiv:2403.17887,

    work page arXiv

  11. [11]

    How do llms use their depth?arXiv preprint arXiv:2510.18871,

    Akshat Gupta, Jay Yeung, Gopala Anumanchipalli, and Anna Ivanova. How do llms use their depth? arXiv preprint arXiv:2510.18871,

    work page arXiv

  12. [12]

    Overthinking the truth: Understanding how language models process false demonstrations.arXiv preprint arXiv:2307.09476,

    Danny Halawi, Jean-Stanislas Denain, and Jacob Steinhardt. Overthinking the truth: Understanding how language models process false demonstrations.arXiv preprint arXiv:2307.09476,

    work page arXiv

  13. [13]

    What affects the effective depth of large language models? arXiv preprint arXiv:2512.14064,

    Yi Hu, Cai Zhou, and Muhan Zhang. What affects the effective depth of large language models? arXiv preprint arXiv:2512.14064,

    work page arXiv

  14. [14]

    arXiv preprint arXiv:2406.19384 , year=

    Vedang Lad, Jin Hwa Lee, Wes Gurnee, and Max Tegmark. The remarkable robustness of llms: Stages of inference?arXiv preprint arXiv:2406.19384,

    work page arXiv

  15. [15]

    Racing thoughts: Explaining contextualization errors in large language models

    Michael A Lepori, Michael Curtis Mozer, and Asma Ghandeharioun. Racing thoughts: Explaining contextualization errors in large language models. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 3020–3036,

    2025

  16. [16]

    Interpreting key mechanisms of factual recall in transformer-based language models

    Ang Lv, Yuhan Chen, Kaiyi Zhang, Yulong Wang, Lifeng Liu, Ji-Rong Wen, Jian Xie, and Rui Yan. Interpreting key mechanisms of factual recall in transformer-based language models.arXiv preprint arXiv:2403.19521,

    work page arXiv

  17. [17]

    The expressive power of transformers with chain of thought, 2024

    William Merrill and Ashish Sabharwal. The expressive power of transformers with chain of thought. arXiv preprint arXiv:2310.07923,

    work page arXiv

  18. [18]

    A little depth goes a long way: The expressive power of log-depth transformers

    William Merrill and Ashish Sabharwal. A little depth goes a long way: The expressive power of log-depth transformers.arXiv preprint arXiv:2503.03961,

    work page arXiv

  19. [19]

    Language models implement simple word2vec- style vector arithmetic

    Jack Merullo, Carsten Eickhoff, and Ellie Pavlick. Language models implement simple word2vec- style vector arithmetic. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pp. 5030–5047,

    2024

  20. [20]

    Attention sinks and compression valleys in llms are two sides of the same coin.arXiv preprint arXiv:2510.06477,

    Accessed: 2026-02-18. Enrique Queipo-de Llano, ´Alvaro Arroyo, Federico Barbero, Xiaowen Dong, Michael Bronstein, Yann LeCun, and Ravid Shwartz-Ziv. Attention sinks and compression valleys in llms are two sides of the same coin.arXiv preprint arXiv:2510.06477,

    work page arXiv 2026

  21. [21]

    Understanding transformer reasoning capabilities via graph algorithms.Advances in Neural Information Processing Systems, 37:78320–78370,

    11 ICLR 2026 Workshop on Logical Reasoning of Large Language Models Clayton Sanford, Bahare Fatemi, Ethan Hall, Anton Tsitsulin, Mehran Kazemi, Jonathan Halcrow, Bryan Perozzi, and Vahab Mirrokni. Understanding transformer reasoning capabilities via graph algorithms.Advances in Neural Information Processing Systems, 37:78320–78370,

    2026

  22. [22]

    Reasoning with latent thoughts: On the power of looped transformers.arXiv preprint arXiv:2502.17416, 2025

    Nikunj Saunshi, Nishanth Dikkala, Zhiyuan Li, Sanjiv Kumar, and Sashank J Reddi. Reasoning with latent thoughts: On the power of looped transformers.arXiv preprint arXiv:2502.17416,

    work page arXiv

  23. [23]

    Clutrr: A diagnostic benchmark for inductive reasoning from text.arXiv preprint arXiv:1908.06177,

    Koustuv Sinha, Shagun Sodhani, Jin Dong, Joelle Pineau, and William L Hamilton. Clutrr: A diagnostic benchmark for inductive reasoning from text.arXiv preprint arXiv:1908.06177,

    work page arXiv 1908

  24. [24]

    Emergent Abilities of Large Language Models

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yo- gatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.arXiv preprint arXiv:2206.07682,

    work page internal anchor Pith review arXiv

  25. [25]

    Transformers: State-of-the-art natural language processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, et al. Transformers: State-of-the-art natural language processing. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, pp. 38–45,

    2020

  26. [26]

    How do transformers learn variable binding in symbolic programs?arXiv preprint arXiv:2505.20896,

    Yiwei Wu, Atticus Geiger, and Rapha ¨el Milli `ere. How do transformers learn variable binding in symbolic programs?arXiv preprint arXiv:2505.20896,

    work page arXiv

  27. [27]

    2023 , archivePrefix=

    Fred Zhang and Neel Nanda. Towards best practices of activation patching in language models: Metrics and methods.arXiv preprint arXiv:2309.16042,

    work page arXiv

  28. [28]

    12 ICLR 2026 Workshop on Logical Reasoning of Large Language Models A IMPLEMENTATION DETAILS All models are accessed through the huggingface transformers library (Wolf et al., 2020); see Table 1 for their links. The logit lens analyses were implemented by accessing the intermediate hidden states and LM-head modules using the functionalities present in the...

    2026

  29. [29]

    Table 1: All pretrained models used in this study, with HuggingFace identifiers, parameter counts, and number of transformer layers

    library. Table 1: All pretrained models used in this study, with HuggingFace identifiers, parameter counts, and number of transformer layers. Family HuggingFace identifier Parameters Layers GPT-2 gpt2117M 12 gpt2-medium345M 24 gpt2-large774M 36 gpt2-xl1.5B 48 Pythia pythia-160m-deduped160M 12 pythia-410m-deduped410M 24 pythia-1.4b-deduped1.4B 24 pythia-2....

    2026

  30. [30]

    The first row is identical to Fig

    (a) Llama3 Family (b) Qwen2 Family Figure B2: Layerwise decoding results for Llama and Qwen2 families 14 ICLR 2026 Workshop on Logical Reasoning of Large Language Models B.2 ADDITIONAL CAUSAL PATCHING RESULTS FOR PRETRAINED MODELS(SIBLINGS-ONLY SETTING) (a) 3-hop: Phi2 (b) 3-hop: Phi4 (c) 3-hop: Phi4-reasoning (d) 5-hop: Phi2 (e) 5-hop: Phi4 (f) 5-hop: Ph...

    2026