Pith. sign in

REVIEW 2 major objections 9 minor 41 references

Control-Diverse Reinforcement Fine-Tuning: Decoupling the Shared Control Bottleneck of RL Post-Training

T0 review · 2 major / 9 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read RL post-training works by changing which components control the reward gain, not merely which ones are activated, and a regularizer that keeps control task-specific improves multi-task accuracy.

desk verdict A clever, careful paper that turns a control-based diagnostic into a trainable regularizer, but the mechanism story hinges on an unvalidated teacher-forced proxy. read the letter →

arxiv 2608.08224 v1 pith:K7PQLHJX submitted 2026-08-08 cs.LG

Binwen Tan , Jingchao Wang , Dengzhe Hou , Lingyu Jiang , Zeyuan Wu , Yunhan Shen , Fangzhou Lin , Kazunori Yamada
show 1 more author
Atsushi Koike
This is my paper · ORCID
classification cs.LG MSC 68T0768T05
keywords reinforcementlearningpost-trainingmechanisticinterpretabilitycontrolcoefficientsharedbottleneckactivation-controlgapGRPOmulti-taskcapabilitylargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that what reinforcement-learning post-training changes inside a language model is which internal components control the reward gain, not merely which components are activated. To make this precise it defines a Post-training Control Coefficient for each component, arranges the coefficients by task family into a control matrix, and measures how much control has collapsed onto a single shared direction across tasks, calling that the Shared Control Bottleneck. Its central empirical claim is that adding a regularizer that lowers this bottleneck to a standard GRPO recipe improves multi-task pass@1 and pass@k on Qwen2.5-7B across mathematics, code, and logic, and that the same signature transfers to Llama-3.2-3B. A sympathetic reader would care because activation-level metrics alone cannot distinguish equally capable models, whereas the control-based diagnostic reproducibly orders them, and it converts a mechanistic diagnosis into a training objective.

What carries the argument

Two objects carry the argument. The Post-training Control Coefficient $C_{m,k} = (1/\ell_m)\,\partial \ell_m(\theta;g)/\partial g_k$ at $g=1$, defined on the reward flux $J_m = \mathbb{E}[\log \pi_\theta(y^\star|x) - \log \pi_0(y^\star|x)]$, is the gate-sensitivity of a task family's gain, read out for all components in a single backward pass by attribution patching. The Shared Control Bottleneck $B_{\mathrm{shared}}(C) = \lambda_{\max}(CC^{\top})/\mathrm{tr}(CC^{\top})$ is the normalized largest eigenvalue of the task-family Gram matrix, measuring concentration of control onto one direction; subtracting it from the activation concentration $B_{\mathrm{shared}}(F)$ gives the Activation–Control Gap. For training, $B_{\mathrm{shared}}$ is replaced by the spectral-moment ratio $R(C) = \mathrm{tr}(G_C^2)/\mathrm{tr}(G_C)^2$, a smooth eigendecomposition-free proxy, and its gradient is reduced to a single backward pass through a central difference in gate space along the closed-form sensitivity direction $W$, which is what makes the regularizer compatible with flash attention and cheap enough for full 7B fine-tuning.

What would settle it

Recompute the control coefficients and the Shared Control Bottleneck with the reward flux replaced by the expected verifier score over sampled completions on the same probe: if the ordering base > GRPO > CD-RFT does not survive, or if pushing the proxy bottleneck down leaves the reward-based bottleneck unchanged while pass@k still rises, the mechanism claim fails. A second check is behavioral: on task families whose control rows are already nearly orthogonal, CD-RFT should show no capability gain; if it does, decoupling is not the active ingredient.

Watch

Extended reading notes

Core claim

The paper's central discovery is that activation and control are separable axes in RL post-training: a component whose output grows in magnitude under fine-tuning need not be one whose small perturbation moves the reward gain. The Post-training Control Coefficient $C_{m,k} = (1/\ell_m)\,\partial \ell_m(\theta;g)/\partial g_k$ evaluated at the nominal gates $g=1$ reads which sublayer components control the reward flux for each task family, and the Shared Control Bottleneck $B_{\mathrm{shared}}(C) = \lambda_{\max}(CC^{\top})/\mathrm{tr}(CC^{\top})$ measures how much of the control energy lies along a single shared direction. On paired recipes that differ only in the regularizer, CD-RFT lowers $B_{\mathrm{shared}}(C)$ relative to GRPO and enlarges the Activation–Control Gap, while the no-KL variant leads on pass@1 and the KL-penalized variant restores the large-$k$ pass@k coverage that reward maximization otherwise degrades. The ordering base > GRPO > CD-RFT in control sharing, and the accompanying capability gains, reproduce on Llama-3.2-3B.

Load-bearing premise

The load-bearing premise is that the reward flux whose control is measured and regularized, the teacher-forced expected log-likelihood margin on a single verified reference completion, is a faithful proxy for the verifier reward that RL post-training actually optimizes; if the components controlling the true reward are different from the ones controlling this margin, both the diagnosis and the regularizer act on the wrong object.

Editorial extensions

If this is right

  • A control-space regularizer becomes a practical lever for multi-task post-training: toggling only the regularizer on top of a matched GRPO recipe improves greedy pass@1 on 8 of 9 benchmarks in the no-KL variant and large-k pass@k coverage on 8 of 9 in the KL-penalized variant.
  • Activation-level mechanistic metrics are insufficient evidence for what post-training changed: of the three activation statistics used by the prior account, one is direction-unstable across equally capable models, while $B_{\mathrm{shared}}(C)$ orders base > GRPO > CD-RFT reproducibly with paired seeds.
  • The known large-k coverage collapse of reward-maximizing RLVR has a control-space remedy: CD-RFT keeps control decoupled while the reward flux is optimized and restores hard-set coverage to base level or above, with the largest gain at the largest k.
  • The regularizer is deployment-compatible: a first-order proxy with measured worst-case per-step overhead of about 8 percent, no second-order graph, and compatibility with flash attention and parameter sharding, running full-parameter 7B fine-tuning on a single GPU.
  • The mechanistic signature and the capability gains transfer across model families and scales, with five base models above the 99th percentile of a random control-matrix reference and Llama-3.2-3B reproducing both the ordering of control sharing and the capability gains.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper establishes a correlation between lower $B_{\mathrm{shared}}(C)$ and higher capability; a natural next experiment, not run here, is to probe whether task pairs whose control rows are already nearly orthogonal gain nothing from CD-RFT, which would test whether decoupling is the mediator rather than a by-product.
  • Because the control coefficient is computed on a teacher-forced margin, the same measurement pipeline could be rerun with the reward flux defined over sampled verifier scores; if the bottleneck ordering survives that redefinition, the mechanism claim would be substantially stronger than the paper's acknowledged proxy.
  • The bottleneck metric is probe- and tokenizer-dependent, so the paper's cross-model comparisons are by ordering only; a normalized or intrinsic-reference version of $B_{\mathrm{shared}}$ would be needed before the metric could serve as a portable benchmark across institutions.
  • If the control view is right, it also predicts an interference diagnostic: tasks whose control rows align along the same direction should interfere most under joint training, so row cosines of the control matrix could predict, before training, which task mixtures would most benefit from CD-RFT.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 9 minor

Summary. This paper proposes a control-based (rather than activation-based) account of what changes inside an LLM during reinforcement-learning post-training. The authors define a reward flux J_m(θ) as the teacher-forced log-likelihood-ratio margin on a verified reference completion (Eq. 1), read out a Post-training Control Coefficient C_{m,k} as the scaled gate-sensitivity of the target log-likelihood (Eq. 4), and summarize cross-task structure through the Shared Control Bottleneck B_shared(C) (Eq. 7) and the Activation–Control Gap (Eq. 8). The main claims are: (i) activation magnitude and control decouple (Lemma 1) and control is generically distributed over gates (Theorem 1); (ii) a regularizer on B_shared, CD-RFT, implemented with a first-order central-difference proxy gradient (Theorem 3), lowers B_shared below a matched GRPO baseline and improves multi-task pass@1/pass@k on Qwen2.5-7B, with the same pattern on Llama-3.2-3B; and (iii) the activation-level 'information complexity' metric is direction-unstable and therefore insufficient. The experiments use same-seed matched pairs, a contamination audit, temperature/checkpoint/target sweeps, and seven-seed paired probe statistics.

Significance. If correct, the paper would add a genuinely new axis—control versus activation—to the mechanistic study of RL post-training, and it is unusual in closing the loop from a mechanistic diagnostic to a training objective with measured capability gains. The strengths are concrete: the experimental protocol is careful (fixed-seed paired comparisons, contamination audit with normalized and near-duplicate matching, temperature and checkpoint sweeps, and a second model family with a pre-fixed target concentration); the first-order proxy-gradient construction is derived in detail and its chain-rule structure is internally coherent; the internal fidelity of the spectral-moment proxy R to B_shared is verified during training (Spearman 0.995, Figure 2b); and the paper repeatedly discloses its own limitations in text (Remark A.2 on the teacher-forced margin, Remark B.2 on the controlled-variable nature of the claim, Appendix F.3 on the row-norm confound).

major comments (2)
  1. [§5.6 (Algorithm 1), §6.3–6.4, Theorem 2 / Theorem C.1] The stress-test concern lands: the central mechanistic claim is about control of the post-training reward gain, but the object actually measured is the gate-sensitivity of a teacher-forced log-likelihood margin on a fixed verified completion y*, not of the expected verifier reward over sampled rollouts that the GRPO/RLVR backbone optimizes. A policy can increase rollout verifier reward (for example, by shifting probability mass onto correct solutions other than y*) while log π_θ(y*|x) decreases; in that regime C_{m,k}, B_shared(C), and the CD-RFT regularizer all describe the wrong sensitivity structure. Remark A.2 concedes that J_m is 'a teacher-forced likelihood margin on one reference target, not the sampling-time expected verifier reward' and that raising it 'need not raise coverage,' but no experiment validates that the margin's gate-sensitivity structure matches the verifier reward's. The conceptual argument in §3.1 that the log-ratio appears in RLVR/DPO/RLHF does not close the gap, because RLVR's actual objective is an advantage-weighted average over sampled rollouts, not the margin on one reference completion. I recommend adding a validation experiment: compute a rollout-based control coefficient (for example, a score-function estimator of ∂_g E[verifier reward] over a fixed rollout set) for GRPO and CD-RFT and compare its Gram structure and B_shared with the margin-based version, or at minimum report the correlation between the two sensitivity vectors. The benchmark claims of Section 6.3 stand independently, but the title-level claim of decoupling 'the shared control bottleneck of RL post-training' is only as strong as this validation; at minimum, the 'reward gain' phrasing should be qualified as margin-based control where it is not the rollout reward.
  2. [§5.6 (Algorithm 1), §6.3–6.4, Theorem 2 / Theorem C.1] The design supports the claim 'adding the regularizer lowers B_shared(C) and capability rises in tandem,' but it does not yet support the causal reading 'lowering B_shared(C) is what improves capability,' because the CD-RFT arm differs from matched GRPO not only in the regularizer value but in the entire inner-loop optimization signal: up to K_max=12 extra gradient steps per training step at η_proj=4×10^-3 on a probe of n=3 sequences per family (9 sequences total). That additional signal, applied at a learning rate 2000× the backbone rate, could plausibly act as a generic inductive bias (implicit flatness, probe-loss smoothing, or a brake on GRPO's trajectory) rather than specifically as control decoupling. The τ-sweep in Appendix I tests the location of the constraint but not this alternative. A compute-matched control is needed: the same inner-loop budget (same τ, K_max, η, and probe) applied to an objective that does not lower B_shared—for example, projection onto random gate directions or onto the activation axis—with both B_shared and the capability numbers reported. The related theoretical claim of Theorem 2 / Theorem C.1 ('irreplaceability of the control axis') is also over-stated as written: the proof rules out a scalarized activation objective Σ_k w_k A_k, but a matrix-level activation regularizer (the same B_shared applied to the F matrix) retains the family dimension and is not addressed by the proof; the operative argument is the activational ceiling B_shared(F)≈99.6% of Section 4.2, which should be made the stated reason rather than the current per-gate scalarization argument.
minor comments (9)
  1. [Table 1 vs Figure F.2] Table 1 and Figure F.2 report inconsistent paired values for the same β=0 comparison (−15.5±6.5, 6/7 in Table 1; −12.3±3.1, 6/7 in Figure F.2); these should be reconciled or the differing computation explained.
  2. [Appendix L] Appendix L states that all CD-RFT arms were trained on RTX PRO 6000 and all GRPO arms on A100, so 'every setting except the regularizer' is not literally true; given bf16 nondeterminism and the few-point margins, the paper should justify why the GPU split cannot affect the paired ordering, or rerun one pair on a single GPU type.
  3. [Abstract, §5.1, Remark C.4 vs Appendix L] The main-text claim of a 'worst-case overhead below eight percent' is contradicted by Appendix L, which reports +11.8% on the median step in the fully-engaged configuration; the abstract and §5.1 should be aligned with the appendix.
  4. [Table 2, Appendix E] Table 2's 'overall pass@k' averages benchmarks evaluated at different k (pass@16, pass@64, pass@256); please state the per-domain k in the caption and consider a common-k summary so the overall column is interpretable as a single sampling budget.
  5. [§5.5, Algorithm 1, Figure F.2] Please state whether the probe used for the τ constraint and inner-loop projection is the same as one of the seven probe seeds over which B_shared(C) is reported, and confirm that the reported drop is maintained on probes not used for training the constraint.
  6. [Theorem 2 / Theorem C.1] Theorem 2's statement should be narrowed as described in Major Comment 2, and the Section 4.2 ceiling argument should be cited there as the operative reason that activation-axis regularizers cannot act.
  7. [Appendix F.2] Appendix F.2 shows that all trained variants (B_shared 69–83%) remain far above the simulated random reference (42.8±3.4), so the 'task-specific control' claim is relative, and the text should clarify that the random reference is not a target for the regularizer.
  8. [§6.5, Table H.1] Appendix H reports a pass@1 win rate of only 3/8 on the logic domain across temperatures; the concluding sentence of the temperature paragraph in §6.5 should be qualified to note that the ordering flips on logic at several temperatures.
  9. [Table 3, Appendix K] Appendix K evaluates the Llama-3.2-3B base zero-shot while trained arms are few-shot; the main-text Table 3 should carry a footnote stating this protocol mismatch explicitly, since the base column is not on the same evaluation protocol as the trained arms.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central capability claims rest on held-out benchmark comparisons against matched GRPO baselines, not on the definitions of the control quantities.

full rationale

The paper's main empirical claims are controlled benchmark comparisons: CD-RFT and its matched GRPO share every setting except the regularizer, and the capability columns (pass@1, pass@k) are measured on nine held-out benchmarks. The quantity B_shared(C) is defined from the control matrix C (Eqs. 4 and 7), and the regularizer in Eq. (9) directly minimizes B_shared; therefore observing that CD-RFT lowers B_shared relative to GRPO is a statement about the objective being optimized, not a fitted prediction or a benchmark-derived constant. No benchmark number is computed from B_shared, so the capability gains cannot be forced by construction. The mechanistic statements (Lemma 1, Theorem 1) are proven from the gating definitions in the paper rather than imported from prior work by the same authors; the cited activation-line work (Zhang et al. 2026) is external and is explicitly challenged, not used as load-bearing support. The acknowledged limitation in Remark A.2, that J_m is a teacher-forced likelihood margin rather than the sampling-time expected verifier reward, is a validity/scope concern about whether the right control object is being measured; it is not a circular reduction in the derivation of the empirical results. No equation in the paper reduces a predicted quantity to a fitted input or to a self-citation chain.

Assumptions & free parameters 5 free parameters · 6 assumptions · 3 invented entities

The central claim rests on many modeling choices: a teacher-forced reward-flux proxy, first-order attribution as control, sublayer gate granularity, probe size n=3, and the regularizer hyperparameters (lambda, tau, epsilon, inner-loop cap). None of these are fitted to the benchmark scores in a hidden way, but several are hand-chosen and the proxy-flux assumption is load-bearing.

free parameters (5)
  • target concentration tau = 70 (Qwen2.5-7B), 50 (Llama-3.2-3B)
    Regularizer target; main value selected from sweep {55,65,70} and affects the pass@1 vs pass@k trade-off.
  • regularizer weight lambda = 1
    Soft penalty coefficient in Eq. (9); no ablation reported for lambda.
  • finite-difference step epsilon = 0.05
    Gate-space perturbation for the central-difference proxy; chosen working point.
  • probe size n = 3 per family
    Sets the estimator Bshared(C); the paper notes the estimator is n-dependent and biased (Prop C.3), and n is set by probe compute budget.
  • inner-loop cap Kmax and step eta = 12, 4e-3
    Near-hard constraint inner loop; chosen working point for the projection.
assumptions (6)
  • domain assumption Smoothness and integrability of the gated log-likelihood (Assumption A.1)
    Needed to commute expectation and differentiation and to justify Taylor and central-difference expansions; holds for SwiGLU/softmax/RMSNorm transformers.
  • domain assumption Nominal gate identity (Assumption A.2)
    Gates at 1 recover the original model, so control quantities are computed near g=1.
  • domain assumption Normalizer non-degeneracy (Assumption A.3)
    Requires |ell_m| >= eps0 so the control coefficient in Eq. (4) is identifiable; families violating it are discarded.
  • ad hoc to paper Teacher-forced log-likelihood ratio proxies the RL reward gain
    Eq. (1) defines J_m as the margin on one verified reference completion, not the actual rollout verifier reward; the paper does not validate this proxy against true reward control.
  • domain assumption First-order attribution captures causal control
    The control coefficient is the linearized attribution-patching effect (Eq. (16)); no intervention experiment confirms that high-control components gate actual reward more than high-activation components.
  • domain assumption Residual-stream superposition and generic weights (Theorem 1)
    Used to argue control support is generically not a single gate; relies on real-analytic activations and measure-zero generic-weight arguments.
invented entities (3)
  • Post-training Control Coefficient C_{m,k}
    purpose: Measures component-level gate sensitivity of the reward flux, claimed to capture control over post-training reward gain.
    Authors-defined normalized gate derivative; no independent falsifiable handle outside the paper's own measurements.
  • Shared Control Bottleneck Bshared(C)
    purpose: Quantifies cross-task concentration of control; serves as both mechanistic diagnostic and regularizer target.
    An invented spectral concentration statistic; its status as the right object rests on the paper's correlation with benchmark gains.
  • Activation-Control Gap ACG
    purpose: Difference between activation sharing and control sharing; claimed to diagnose loss of task-specific control.
    Derived from two authors-defined matrices; no external evidence that the gap itself is causally meaningful.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Control-Diverse Reinforcement Fine-Tuning: Decoupling the Shared Control Bottleneck of RL Post-Training." pith.science (2026). https://pith.science/paper/K7PQLHJX

@misc{pith2026260808224,
  author       = {Pith},
  title        = {Pith review of: Control-Diverse Reinforcement Fine-Tuning: Decoupling the Shared Control Bottleneck of RL Post-Training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7PQLHJX}},
  note         = {Machine review of arXiv:2608.08224}
}
read the original abstract

Reinforcement learning post-training unlocks complex reasoning in LLMs. Yet benchmark scores reveal only whether a model improved, not what changed inside it, nor how it splits finite capability across tasks. A representative interpretability line attributes the success of RL fine-tuning to stronger and more diverse circuit activation. We challenge this activation-centered account by separating activation from control: an activated circuit need not control the post-training reward gain. Adapting Metabolic Control Analysis, we define the Post-training Control Coefficient to measure component control over reward gain and arrange these coefficients by task family into a control matrix, paired with an activation-magnitude matrix. We call cross-task control concentration the Shared Control Bottleneck and the difference between activation and control concentration the Activation-Control Gap. This reveals that highly shared activations can coexist with task-specific control, while a small gap indicates that control has collapsed onto a shared direction and lost task specificity. To reduce this collapse, we regularize the post-training loss with the Shared Control Bottleneck and propose Control-Diverse Reinforcement Fine-Tuning (CD-RFT). The exact regularizer gradient requires second-order automatic differentiation incompatible with flash attention, so we derive a first-order proxy with worst-case overhead below eight percent. On Qwen2.5-7B, CD-RFT achieves the largest control decoupling and improves multi-task capability over matched GRPO across mathematics, code, and logic. The no-KL variant leads on pass@1, and the KL-penalized variant leads on large-k pass@k coverage that KL otherwise degrades. Together, these results show that the Shared Control Bottleneck is both a mechanistic diagnostic and a training regularizer, and that control decoupling and capability gains transfer to Llama-3.2-3B.

Figures

Figures reproduced from arXiv: 2608.08224 by the authors.

Figure 1
Figure 1. The control view at a glance. Top: task-wise reward flux Jm (Eq. (1)). Middle: attribution patching (Syed, Rager, and Conmy 2024) through a differentiable gate on each component yields the control coefficient Cm,k (Eq. (4)); A and B show that activation does not imply control. Bottom: CD-RFT keeps control task-specific across families, lowering Bshared(C) and raising pass@1/pass@k. Benchmark scores answer only wheth… view at source ↗
Figure 2
Figure 2. Distributedness of control and proxy fidelity. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Information complexity (edge-distribution entropy) [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 21 canonical work pages

  1. [1]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Training Language Models to Follow Instructions with Human Feedback , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  2. [2]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Direct Preference Optimization: Your Language Model is Secretly a Reward Model , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  3. [3]

    arXiv preprint arXiv:2402.03300 , year=

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models , author=. arXiv preprint arXiv:2402.03300 , year=

    arXiv

  4. [4]

    arXiv preprint arXiv:2501.12948 , year=

    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning , author=. arXiv preprint arXiv:2501.12948 , year=

    arXiv

  5. [5]

    arXiv preprint arXiv:2503.14476 , year=

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale , author=. arXiv preprint arXiv:2503.14476 , year=

    arXiv

  6. [6]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  7. [7]

    and Liu, Alisa and Dziri, Nouha and Lyu, Shane and others , journal=

    Lambert, Nathan and Morrison, Jacob and Pyatkin, Valentina and Huang, Shengyi and Ivison, Hamish and Brahman, Faeze and Miranda, Lester James V. and Liu, Alisa and Dziri, Nouha and Lyu, Shane and others , journal=. T

    work page

  8. [8]

    2025 , howpublished=

    Code-R1: Reproducing R1 for Code with Reliable Rewards , author=. 2025 , howpublished=

    work page 2025

Show all 41 references
  1. [9]

    2025 , howpublished=

    DeepScaleR: Surpassing O1-Preview with a 1.5B Model by Scaling RL , author=. 2025 , howpublished=

    2025

  2. [10]

    arXiv preprint arXiv:2506.14965 , year=

    Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective , author=. arXiv preprint arXiv:2506.14965 , year=

    arXiv

  3. [11]

    arXiv preprint arXiv:2602.05547 , year=

    Multi-Task GRPO: Reliable LLM Reasoning Across Tasks , author=. arXiv preprint arXiv:2602.05547 , year=

    arXiv

  4. [12]

    Proceedings of the 7th BlackboxNLP Workshop (EMNLP) , year=

    Attribution Patching Outperforms Automated Circuit Discovery , author=. Proceedings of the 7th BlackboxNLP Workshop (EMNLP) , year=

  5. [13]

    2021 , howpublished=

    A Mathematical Framework for Transformer Circuits , author=. 2021 , howpublished=

    2021

  6. [14]

    2023 , howpublished=

    Attribution Patching: Activation Patching At Industrial Scale , author=. 2023 , howpublished=

    2023

  7. [15]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Towards Automated Circuit Discovery for Mechanistic Interpretability , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  8. [16]

    Interpretability in the Wild: a Circuit for Indirect Object Identification in

    Wang, Kevin and Variengien, Alexandre and Conmy, Arthur and Shlegeris, Buck and Steinhardt, Jacob , booktitle=. Interpretability in the Wild: a Circuit for Indirect Object Identification in

  9. [17]

    arXiv preprint arXiv:2403.00745 , year=

    Kram. arXiv preprint arXiv:2403.00745 , year=

    arXiv

  10. [18]

    International Conference on Machine Learning (ICML) , year=

    Axiomatic Attribution for Deep Networks , author=. International Conference on Machine Learning (ICML) , year=

  11. [19]

    Conference on Language Modeling (COLM) , year=

    Have Faith in Faithfulness: Going Beyond Circuit Overlap When Finding Model Mechanisms , author=. Conference on Language Modeling (COLM) , year=

  12. [20]

    International Conference on Learning Representations (ICLR) , year=

    Fine-Tuning Enhances Existing Mechanisms: A Case Study on Entity Tracking , author=. International Conference on Learning Representations (ICLR) , year=

  13. [21]

    International Conference on Learning Representations (ICLR) , year=

    Mechanistically Analyzing the Effects of Fine-Tuning on Procedurally Defined Tasks , author=. International Conference on Learning Representations (ICLR) , year=

  14. [22]

    International Conference on Learning Representations (ICLR) , year=

    Reinforcement Learning Fine-Tuning Enhances Activation Intensity and Diversity in the Internal Circuitry of LLMs , author=. International Conference on Learning Representations (ICLR) , year=

  15. [23]

    Biochemical Journal , volume=

    Metabolic Control Analysis: A Survey of Its Theoretical and Experimental Development , author=. Biochemical Journal , volume=

  16. [24]

    Symposia of the Society for Experimental Biology , volume=

    The Control of Flux , author=. Symposia of the Society for Experimental Biology , volume=

  17. [25]

    European Journal of Biochemistry , volume=

    A Linear Steady-State Treatment of Enzymatic Chains: General Properties, Control and Effector Strength , author=. European Journal of Biochemistry , volume=

  18. [26]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  19. [27]

    International Conference on Learning Representations (ICLR) , year=

    Semi-Supervised Classification with Graph Convolutional Networks , author=. International Conference on Learning Representations (ICLR) , year=

  20. [28]

    , journal=

    Davis, Chandler and Kahan, William M. , journal=. The Rotation of Eigenvectors by a Perturbation

  21. [29]

    Matrix Differential Calculus with Applications in Statistics and Econometrics , author=

  22. [30]

    Inequalities: Theory of Majorization and Its Applications , author=

  23. [31]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating Large Language Models Trained on Code , author=. arXiv preprint arXiv:2107.03374 , year=

    arXiv

  24. [32]

    Measuring Mathematical Problem Solving with the

    Hendrycks, Dan and Burns, Collin and Kadavath, Saurav and Arora, Akul and Basart, Steven and Tang, Eric and Song, Dawn and Steinhardt, Jacob , booktitle=. Measuring Mathematical Problem Solving with the

  25. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Solving Quantitative Reasoning Problems with Language Models , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

  26. [34]

    arXiv preprint arXiv:2108.07732 , year=

    Program Synthesis with Large Language Models , author=. arXiv preprint arXiv:2108.07732 , year=

    arXiv

  27. [35]

    Is Your Code Generated by

    Liu, Jiawei and Xia, Chunqiu Steven and Wang, Yuyao and Zhang, Lingming , booktitle=. Is Your Code Generated by

  28. [36]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal Policy Optimization Algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    arXiv

  29. [37]

    and Ermon, Stefano and Rudra, Atri and R

    Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R. Advances in Neural Information Processing Systems (NeurIPS) , year=

  30. [38]

    Rajbhandari, Samyam and Rasley, Jeff and Ruwase, Olatunji and He, Yuxiong , booktitle=

  31. [39]

    Zhao, Yanli and Gu, Andrew and Varma, Rohan and Luo, Liang and Huang, Chien-Chin and Xu, Min and Wright, Less and Shojanazeri, Hamid and Ott, Myle and Shleifer, Sam and others , journal=

  32. [40]

    arXiv preprint arXiv:2412.15115 , year=

    Qwen2.5 Technical Report , author=. arXiv preprint arXiv:2412.15115 , year=

    arXiv

  33. [41]

    arXiv preprint arXiv:2407.21783 , year=

    The Llama 3 Herd of Models , author=. arXiv preprint arXiv:2407.21783 , year=

    arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.