Pith. sign in

REVIEW 4 major objections 1 cited by

Conflicts in constrained model steering can be fixed by orthogonalizing the spectral subspaces of the two objectives only where they clash.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 10:40 UTC

load-bearing objection SIFT is a usable spectral fix for primary–constraint conflicts in LLM post-training: the merging→Muon bridge is real, the four-task gains hold up, and the load-bearing risk is hand-tuned τ/K localization rather than a broken idea. the 4 major comments →

arxiv 2604.04229 v1 submitted 2026-04-05 cs.MM cs.AIcs.CVcs.SD

Hierarchical Semantic Correlation-Aware Masked Autoencoder for Unsupervised Audio-Visual Representation Learning

Donghuo Zeng , Hao Niu , Masato Taya This is my paper
classification cs.MM cs.AIcs.CVcs.SD
keywords subspace controlspectral optimizationMuonconstrained model steeringmachine unlearningsafety alignmentgradient orthogonalizationtask interference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

When a pretrained model is steered to satisfy a constraint (unlearning, safety, speech generation, anti-hallucination) while keeping a primary objective (utility, text, non-hallucinated content), the two losses often push parameters in opposite directions at specific layers and training steps. The paper shows that this conflict is the same spectral interference that appears when one-shot task vectors are merged with non-orthogonal singular bases. It resolves the interference by taking the top singular directions of the primary and constraint momentum matrices, concatenating them, and applying Muon’s matrix-sign orthogonalization—exactly the whitening step that makes merged subspaces interference-free. The fix is applied only when gradient cosine similarity falls below a threshold, so most of training stays ordinary Muon. Across four steering tasks the method improves the primary–constraint trade-off relative to AdamW, plain Muon, gradient-projection bi-level optimizers, and one-shot spectral edits, while retaining useful descent information that pure projection discards.

Core claim

Primary–constraint conflicts in model steering are spectral interference arising from non-orthogonal task subspaces. Constructing an interference-free subspace from the top-K singular components of both momentum matrices and applying Muon-style matrix-sign orthogonalization only at the layers and steps where cosine similarity is sufficiently negative yields controllable updates that jointly improve both objectives without discarding useful shared descent directions.

What carries the argument

SIFT (spectral interference-free training): when cosine similarity of primary and constraint gradients falls below a threshold, form the concatenated top-K left/right singular subspaces of the two momentum matrices, orthogonalize them with the matrix sign function, and update with that outer product; otherwise fall back to ordinary Muon.

Load-bearing premise

A hand-chosen cosine-similarity threshold between the two gradients is assumed to be a reliable detector of harmful spectral interference, and orthogonalizing only the top-K singular directions of those momentum matrices is assumed to remove the conflict without throwing away useful shared signal.

What would settle it

On any of the four tasks, if activating SIFT on every layer and step (or never) matches or beats the localized schedule, or if random subspaces of the same size outperform the top-K primary-plus-constraint subspaces, the localization-and-spectral-interference account fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Unlearning can suppress more hazardous knowledge while keeping utility closer to the base model than projection-based bi-level methods.
  • Safety alignment can raise refusal metrics and utility benchmarks at the same time rather than trading one for the other.
  • Cross-modal speech–text fine-tuning can lift audio-input accuracy without erasing text capability.
  • Word-level hallucination can be reduced while preserving non-hallucinated content and downstream scores.
  • Interference is sparse and task-specific (higher layers for unlearning; middle layers early for safety; QKV for speech and hallucination), so control need not be applied globally.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same sparse localization-plus-orthogonalization recipe may extend to multi-constraint steering if pairwise interference remains sparse.
  • If post-training momentum stays low-rank, truncated SVD could close much of SIFT’s runtime gap versus plain Muon.
  • Constrained pre-training might benefit from the same spectral control once the cosine detector is retuned for denser, earlier conflicts.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 0 minor

Summary. The paper formulates model steering as constrained optimization with a primary objective f and a constraint objective g (hard or soft), and argues that poor primary–constraint trade-offs arise from structured, localized gradient misalignment. Motivated by spectral interference in one-shot task-vector merging and by Muon’s matrix-sign (msign) gradient orthogonalization, it proposes SIFT: at steps and layers where cosine similarity τ between primary and constraint gradients/momenta falls below a threshold ε, it concatenates the top-K singular subspaces of the two momenta, applies msign to obtain interference-free bases, and updates with the resulting orthogonalized direction; otherwise it falls back to standard Muon. The method is evaluated on four post-training tasks—LLM unlearning (RMU/WMDP), safety alignment (DPO), text-to-speech adaptation (GLM-4-Voice), and hallucination mitigation (RAGTruth)—against AdamW, Muon, POME, and BLUR, with reported gains on primary–constraint trade-offs, sensitivity analyses on K, and spatial–temporal localization heatmaps.

Significance. If the empirical trade-off improvements hold under broader scrutiny, SIFT is a useful practical contribution: it reframes constrained post-training as controllable spectral optimization, gives a concrete link between model-merging orthogonalization and Muon, and shows consistent gains over both control-free optimizers and control-based baselines (BLUR, POME) on four distinct steering problems. Strengths include multi-task evaluation with means±std over trials, external benchmarks (ES/MCQ, safety suites, MMLU/GSM8K, cross-modal accuracy, hallucination rate), sensitivity plots for subspace rank K, localization visualizations, and a public code link. The work is primarily empirical; there is no convergence or interference-elimination theorem, so significance rests on the robustness and generality of the reported trade-offs rather than on a closed-form guarantee.

major comments (4)
  1. §3–§4 and Algorithm A1: The load-bearing premise is that scalar cosine similarity τ (thresholded by a hand-chosen ε<0) is a sufficient detector of the spectral interference that SIFT is designed to remove, and that msign on the concatenated top-K bases of Mf and Mg (Eqs. 8–10) removes harmful interference while retaining useful shared descent information. The paper supplies only a one-shot merging analogy and final-task metrics. Please add (i) a direct diagnostic that the SIFT update reduces a measurable interference quantity (e.g., non-orthogonality of task subspaces or negative contribution of the discarded component as in Fig. 3) at the intervened steps/layers, and (ii) an ablation of always-on SIFT vs τ-triggered localization vs random localization, so that the necessity of the τ policy is tested rather than assumed.
  2. §5.1 and Tables 3–6: K and ε are chosen case-by-case (K=128/192 for unlearning/safety; full rank for TTS and hallucination; ε from −0.1 to −0.8). Gains and the “sparse localization” story therefore depend on per-task tuning. Please report a fixed or cross-validated selection protocol (or a small grid with held-out criteria) and show that the ranking vs BLUR/Muon/POME is stable under modest (K,ε) misspecification; otherwise the central claim of a general subspace-control method is overstated relative to a heavily tuned intervention.
  3. §4 (steps a–d) and comparison to BLUR (Eq. 3, Fig. 3): Fig. 3 shows that the component BLUR discards still reduces utility loss, motivating retention of both subspaces. SIFT claims to retain both while eliminating interference via msign, but there is no intermediate analysis (e.g., primary and constraint loss trajectories, or alignment of the SIFT direction with ∇f and ∇g) demonstrating that the update is not simply a softer form of projection or a reweighted Muon step. A short analysis or controlled synthetic experiment would make the mechanism claim load-bearing rather than post-hoc.
  4. Runtime and scalability (Tables 3–6): SIFT roughly doubles runtime vs Muon (e.g., 20.2 vs 11.8 min unlearning; 215.6 vs 191.2 safety; 26.4 vs 12.0 hallucination), attributed to SVDs for top-K extraction. Given that localization is sparse (Fig. 4), please quantify the fraction of steps/layers where SIFT fires and the incremental cost of those SVDs, and discuss whether cheaper approximations (randomized SVD, fewer Newton–Schulz steps, or caching) preserve the reported gains. Without this, the practicality claim for 7B-scale post-training is incomplete.

Circularity Check

0 steps flagged

No significant circularity: SIFT is an empirical spectral-control method whose gains are measured on external benchmarks, not quantities defined by its own update rule or self-citation chain.

full rationale

The manuscript (body is SIFT despite HSC-MAE metadata) proposes a localized Muon-style update that orthogonalizes concatenated top-K momentum subspaces of primary and constraint objectives when cosine similarity τ falls below a threshold ε (Eqs. 8–10, Alg. A1). This is motivated by a one-shot model-merging analogy (Eqs. 4–5) and the matrix-sign step of Muon (Eqs. 6–7), but the analogy is not claimed as a uniqueness theorem or first-principles derivation that forces the experimental outcomes. All reported results (Tables 3–6) are external metrics—ES/MCQ unlearning scores, MMLU/GSM8K/IFEval utility, StrongReject/JBB safety, cross-modal accuracy, RAGTruth hallucination rate—none of which are defined by the SIFT update, the fitted (K, ε) pair, or the τ detector. Hyperparameters are tuned per task, yet the evaluation quantities remain independent of those fits. Self-citations (BLUR, RMU, prior OPTML unlearning work) supply baselines and problem formulation only; they do not underwrite the measured gains or close a definitional loop. No step reduces a claimed prediction to its own inputs by construction. Score 1 reflects only the ordinary presence of author-overlapping baseline citations that are not load-bearing for the central claim.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

SIFT is an optimization method paper. Its central claim rests on a small set of modeling choices (τ as interference detector; top-K spectral merge + msign as the fix; sparse localization) plus standard ML training assumptions, not on new physical entities. Free parameters K and ε are load-bearing for reported gains and are selected per application. No machine-checked proofs; evidence is empirical tables.

free parameters (4)
  • subspace dimension K = 128 / 192 / full (task-dependent)
    Number of top singular components kept from each of Mf and Mg before concatenation (Eq. 8). Set to 128 (unlearning), 192 (safety), full rank (TTS, hallucination). Sensitivity plots show performance peaks at specific K.
  • misalignment threshold ε = −0.1 to −0.8 by task
    Triggers SIFT vs plain Muon when τ < ε (Algorithm A1). Chosen case-by-case: −0.1 unlearning/safety; −0.6/−0.4/−0.5 TTS datasets; −0.8 hallucination.
  • regularization / multi-objective weight λ (soft form)
    Soft-regularized formulation (Eq. 2) balances primary f and constraint g; exact schedules are application-inherited and affect the trade-off surface SIFT is judged on.
  • Muon / SIFT step sizes η_t and momentum hyperparameters
    Standard optimizer knobs; not fully tabulated in the main text but required to reproduce Tables 3–6.
axioms (5)
  • domain assumption Spectral non-orthogonality of concatenated primary and constraint task/momentum subspaces is a primary structural cause of objective–constraint conflict in constrained steering.
    Imported from model-merging literature (Gargiulo et al., Marczak et al.) and used as the design premise in §4.
  • ad hoc to paper The matrix sign function (msign / Newton–Schulz) applied to the concatenated top-K bases yields an interference-free descent direction that retains useful components of both objectives.
    Core constructive claim of SIFT (Eqs. 7–10); motivated by analogy to whitening/Procrustes merging, not proved for the iterative constrained case.
  • ad hoc to paper Cosine similarity τ between ∇f and ∇g (or their momenta), thresholded by ε, reliably localizes steps and layers where spectral control is needed.
    Localization policy in §3–§4 and Algorithm A1; empirical heatmaps support sparsity but do not prove optimality of τ.
  • domain assumption Momentum matrices for these post-training tasks are often low-rank enough that small K captures most energy.
    Used to justify K≪d for unlearning/safety (Fig. 5, Fig. A2); not assumed for TTS/hallucination.
  • standard math Standard compact SVD / Newton–Schulz linear algebra and Muon spectral-norm steepest-descent interpretation.
    Background for Eqs. 5–7; cited via Bernstein & Newhouse, Jordan et al.
invented entities (2)
  • SIFT update (interference-free spectral subspace from dual-momentum top-K + msign) no independent evidence
    purpose: Provide a controllable descent direction that jointly optimizes primary and constraint objectives without discarding conflicting components as in gradient projection.
    Defined in §4 (a)–(d) and Eq. 10; existence and benefit are justified empirically, not by an independent external measurement.
  • τ-triggered spatial-temporal localization policy for spectral control no independent evidence
    purpose: Apply expensive subspace control only where misalignment is detected, yielding sparse intervention patterns.
    Algorithm A1 and Figs. 2, 4, 6, A1, A4; policy is paper-specific.

pith-pipeline@v1.1.0-grok45 · 24196 in / 3743 out tokens · 43831 ms · 2026-07-13T10:40:38.391416+00:00 · methodology

0 comments
read the original abstract

Learning aligned multimodal embeddings from weakly paired, label-free corpora is challenging: pipelines often provide only pre-extracted features, clips contain multiple events, and spurious co-occurrences. We propose HSC-MAE (Hierarchical Semantic Correlation-Aware Masked Autoencoder), a dual-path teacher-student framework that enforces semantic consistency across three complementary levels of representation - from coarse to fine: (i) global-level canonical-geometry correlation via DCCA, which aligns audio and visual embeddings within a shared modality-invariant subspace; (ii) local-level neighborhood-semantics correlation via teacher-mined soft top-k affinities, which preserves multi-positive relational structure among semantically similar instances; and (iii) sample-level conditional-sufficiency correlation via masked autoencoding, which ensures individual embeddings retain discriminative semantic content under partial observation. Concretely, a student MAE path is trained with masked feature reconstruction and affinity-weighted soft top-k InfoNCE; an EMA teacher operating on unmasked inputs via the CCA path supplies stable canonical geometry and soft positives. Learnable multi-task weights reconcile competing objectives, and an optional distillation loss transfers teacher geometry into the student. Experiments on AVE and VEGAS demonstrate substantial mAP improvements over strong unsupervised baselines, validating that HSC-MAE yields robust and well-structured audio-visual representations.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. SW-$A^2$-Bench: Benchmarking Autonomous Software Agent Generation for Agentic Web

    cs.MA 2026-04 unverdicted novelty 6.0

    SW-A²-Bench is presented as the first benchmark for generating faithful, interoperable software agents from code repositories via coding agents.

Reference graph

Works this paper leans on

33 extracted references · 19 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Old optimizer, new norm: An anthology.arXiv preprint arXiv:2409.20325,

    Jeremy Bernstein and Laker Newhouse. Old optimizer, new norm: An anthology.arXiv preprint arXiv:2409.20325,

    Pith/arXiv arXiv

  2. [2]

    Muon optimizes under spectral norm constraints.arXiv preprint arXiv:2506.15054,

    Lizhang Chen, Jonathan Li, and Qiang Liu. Muon optimizes under spectral norm constraints.arXiv preprint arXiv:2506.15054,

    arXiv

  3. [3]

    V oicebench: Benchmarking llm-based voice assistants.arXiv preprint arXiv:2410.17196,

    Yiming Chen, Xianghu Yue, Chen Zhang, Xiaoxue Gao, Robby T Tan, and Haizhou Li. V oicebench: Benchmarking llm-based voice assistants.arXiv preprint arXiv:2410.17196,

    Pith/arXiv arXiv

  4. [4]

    Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. Training verifiers to solve math word problems.arXiv preprint arXiv:2110.14168,

    Pith/arXiv arXiv

  5. [5]

    Closing the gap between text and speech understanding in llms.arXiv preprint arXiv:2510.13632,

    Santiago Cuervo, Skyler Seto, Maureen de Seyssel, Richard He Bai, Zijin Gu, Tatiana Likhomanenko, Navdeep Jaitly, and Zakaria Aldeneh. Closing the gap between text and speech understanding in llms.arXiv preprint arXiv:2510.13632,

    arXiv

  6. [6]

    Computer software

    URL �������������������������������������� . Computer software. Chongyu Fan, Changsheng Wang, Yancheng Huang, Soumyadeep Pal, and Sijia Liu. Llm unlearning under the microscope: A full-stack view on methods and metrics.arXiv preprint arXiv:2510.07626,

    arXiv

  7. [7]

    Parameter-efficient fine-tuning for large models: A comprehensive survey.arXiv preprint arXiv:2403.14608,

    Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey.arXiv preprint arXiv:2403.14608,

    Pith/arXiv arXiv

  8. [8]

    Low-rank orthogonalization for large-scale matrix optimization with applications to foundation model training.arXiv preprint arXiv:2509.11983,

    Chuan He, Zhanwang Deng, and Zhaosong Lu. Low-rank orthogonalization for large-scale matrix optimization with applications to foundation model training.arXiv preprint arXiv:2509.11983,

    Pith/arXiv arXiv

  9. [9]

    Safety tax: Safety alignment makes your large reasoning models less reasonable, 2025.URL https://arxiv

    Tiansheng Huang, Sihao Hu, Fatih Ilhan, Selim Furkan Tekin, Zachary Yahn, Yichang Xu, and Ling Liu. Safety tax: Safety alignment makes your large reasoning models less reasonable, 2025.URL https://arxiv. org/abs/2503.00555. Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. Edi...

    Pith/arXiv arXiv 2025

  10. [10]

    Pku-saferlhf: Towards multi-level safety alignment for llms with human preference.arXiv preprint arXiv:2406.15513,

    Jiaming Ji, Donghai Hong, Borong Zhang, Boyuan Chen, Josef Dai, Boren Zheng, Tianyi Qiu, Boxun Li, and Yaodong Yang. Pku-saferlhf: Towards multi-level safety alignment for llms with human preference.arXiv preprint arXiv:2406.15513,

    Pith/arXiv arXiv

  11. [11]

    Beyond sft: Reinforcement learning for safer large reasoning models with better reasoning ability.arXiv preprint arXiv:2512.01848,

    Jinghan Jia, Nathalie Baracaldo, and Sijia Liu. Beyond sft: Reinforcement learning for safer large reasoning models with better reasoning ability.arXiv preprint arXiv:2512.01848,

    arXiv

  12. [12]

    Dmitry Kovalev

    URL������������������������������������� �����. Dmitry Kovalev. Understanding gradient orthogonalization for deep learning via non-euclidean trust-region optimization. arXiv preprint arXiv:2503.12645,

    Pith/arXiv arXiv

  13. [13]

    The wmdp benchmark: Measuring and reducing malicious use with unlearning.arXiv preprint arXiv:2403.03218,

    Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, et al. The wmdp benchmark: Measuring and reducing malicious use with unlearning.arXiv preprint arXiv:2403.03218,

    Pith/arXiv arXiv

  14. [14]

    Mitigating the alignment tax of rlhf

    Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jianmeng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, et al. Mitigating the alignment tax of rlhf. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pp. 580–606,

    2024

  15. [15]

    Muon is scalable for llm training.arXiv preprint arXiv:2502.16982, 2025a

    Jingyuan Liu, Jianlin Su, Xingcheng Yao, Zhejun Jiang, Guokun Lai, Yulun Du, Yidao Qin, Weixin Xu, Enzhe Lu, Junjie Yan, et al. Muon is scalable for llm training.arXiv preprint arXiv:2502.16982, 2025a. Sijia Liu, Yuanshun Yao, Jinghan Jia, Stephen Casper, Nathalie Baracaldo, Peter Hase, Yuguang Yao, Chris Yuhao Liu, Xiaojun Xu, Hang Li, et al. Rethinking ...

    Pith/arXiv arXiv

  16. [16]

    Swan: Sgd with normalization and whitening enables stateless llm training.arXiv preprint arXiv:2412.13148,

    Chao Ma, Wenbo Gong, Meyer Scetbon, and Edward Meeds. Swan: Sgd with normalization and whitening enables stateless llm training.arXiv preprint arXiv:2412.13148,

    Pith/arXiv arXiv

  17. [17]

    Preconditioning benefits of spectral orthogonalization in muon

    Jianhao Ma, Yu Huang, Yuejie Chi, and Yuxin Chen. Preconditioning benefits of spectral orthogonalization in muon. arXiv preprint arXiv:2601.13474,

    arXiv

  18. [18]

    No task left behind: Isotropic model merging with common and task-specific subspaces.arXiv preprint arXiv:2502.04959,

    Daniel Marczak, Simone Magistri, Sebastian Cygert, Bartłomiej Twardowski, Andrew D Bagdanov, and Joost Van De Weijer. No task left behind: Isotropic model merging with common and task-specific subspaces.arXiv preprint arXiv:2502.04959,

    Pith/arXiv arXiv

  19. [19]

    Pointer sentinel mixture models.arXiv preprint arXiv:1609.07843,

    Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models.arXiv preprint arXiv:1609.07843,

    Pith/arXiv arXiv

  20. [20]

    Direct preference optimization: Your language model is secretly a reward model.arXiv preprint arXiv:2305.18290,

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.arXiv preprint arXiv:2305.18290,

    Pith/arXiv arXiv

  21. [21]

    Blur: A bi-level optimization approach for llm unlearning.arXiv preprint arXiv:2506.08164,

    Hadi Reisizadeh, Jinghan Jia, Zhiqi Bu, Bhanukiran Vinzamuri, Anil Ramakrishna, Kai-Wei Chang, V olkan Cevher, Sijia Liu, and Mingyi Hong. Blur: A bi-level optimization approach for llm unlearning.arXiv preprint arXiv:2506.08164,

    arXiv

  22. [22]

    Gluon: Making muon & scion great again!(bridging theory and practice of lmo-based optimizers for llms).arXiv preprint arXiv:2505.13416,

    Artem Riabinin, Egor Shulgin, Kaja Gruntkowska, and Peter Richtárik. Gluon: Making muon & scion great again!(bridging theory and practice of lmo-based optimizers for llms).arXiv preprint arXiv:2505.13416,

    Pith/arXiv arXiv

  23. [23]

    Muse: Machine unlearning six-way evaluation for language models.arXiv preprint arXiv:2407.06460,

    Weijia Shi, Jaechan Lee, Yangsibo Huang, Sadhika Malladi, Jieyu Zhao, Ari Holtzman, Daogao Liu, Luke Zettlemoyer, Noah A Smith, and Chiyuan Zhang. Muse: Machine unlearning six-way evaluation for language models.arXiv preprint arXiv:2407.06460,

    Pith/arXiv arXiv

  24. [24]

    Position: Capability control should be a separate goal from alignment.arXiv preprint arXiv:2602.05164,

    Shoaib Ahmed Siddiqui, Eleni Triantafillou, David Krueger, and Adrian Weller. Position: Capability control should be a separate goal from alignment.arXiv preprint arXiv:2602.05164,

    arXiv

  25. [25]

    Steering llm reasoning through bias-only adaptation

    Viacheslav Sinii, Alexey Gorbatovski, Artem Cherepanov, Boris Shaposhnikov, Nikita Balagansky, and Daniil Gavrilov. Steering llm reasoning through bias-only adaptation. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pp. 9213–9222,

    2025

  26. [26]

    Zijun Wang, Haoqin Tu, Yuhan Wang, Juncheng Wu, Jieru Mei, Brian R

    In the Proceedings of ICLR. Zijun Wang, Haoqin Tu, Yuhan Wang, Juncheng Wu, Jieru Mei, Brian R. Bartoldson, Bhavya Kailkhura, and Cihang Xie. Star-1: Safer alignment of reasoning llms with 1k data.arXiv preprint arXiv:2504.01903,

    arXiv

  27. [27]

    Low-rank adaptation for foundation models: A comprehensive review.arXiv preprint arXiv:2501.00365,

    Menglin Yang, Jialin Chen, Jinkai Tao, Yifei Zhang, Jiahong Liu, Jiasheng Zhang, Qiyao Ma, Harshit Verma, Regina Zhang, Min Zhou, et al. Low-rank adaptation for foundation models: A comprehensive review.arXiv preprint arXiv:2501.00365,

    arXiv

  28. [28]

    Merging beyond: Streaming llm updates via activation-guided rotations.arXiv preprint arXiv:2602.03237,

    Yuxuan Yao, Haonan Sheng, Qingsong Lv, Han Wu, Shuqi Liu, Zehua Liu, Zengyan Liu, Jiahui Gao, Haochen Tan, Xiaojin Fu, et al. Merging beyond: Streaming llm updates via activation-guided rotations.arXiv preprint arXiv:2602.03237,

    arXiv

  29. [29]

    A closer look at machine unlearning for large language models.arXiv preprint arXiv:2410.08109,

    Xiaojian Yuan, Tianyu Pang, Chao Du, Kejiang Chen, Weiming Zhang, and Min Lin. A closer look at machine unlearning for large language models.arXiv preprint arXiv:2410.08109,

    Pith/arXiv arXiv

  30. [30]

    Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot.arXiv preprint arXiv:2412.02612, 2024a

    Aohan Zeng, Zhengxiao Du, Mingdao Liu, Kedong Wang, Shengmin Jiang, Lei Zhao, Yuxiao Dong, and Jie Tang. Glm-4-voice: Towards intelligent and human-like end-to-end spoken chatbot.arXiv preprint arXiv:2412.02612, 2024a. Aohan Zeng, Zhengxiao Du, Mingdao Liu, Lei Zhang, Shengmin Jiang, Yuxiao Dong, and Jie Tang. Scaling speech-text pre-training with synthet...

    Pith/arXiv arXiv 2024

  31. [31]

    Negative preference optimization: From catastrophic collapse to effective unlearning.arXiv preprint arXiv:2404.05868, 2024b

    Ruiqi Zhang, Licong Lin, Yu Bai, and Song Mei. Negative preference optimization: From catastrophic collapse to effective unlearning.arXiv preprint arXiv:2404.05868, 2024b. Xuanqi Zhang, Haoyang Shang, and Xiaoxiao Li. Gss: Gated subspace steering for selective memorization mitigation in llms.arXiv preprint arXiv:2602.08901,

    Pith/arXiv arXiv

  32. [32]

    Hanqing Zhu, Zhenyu Zhang, Hanxian Huang, DiJia Su, Zechun Liu, Jiawei Zhao, Igor Fedorov, Hamed Pirsiavash, Zhizhou Sha, Jinwon Lee, et al

    URL ��������������������������� �����. Hanqing Zhu, Zhenyu Zhang, Hanxian Huang, DiJia Su, Zechun Liu, Jiawei Zhao, Igor Fedorov, Hamed Pirsiavash, Zhizhou Sha, Jinwon Lee, et al. The path not taken: Rlvr provably learns off the principals.arXiv preprint arXiv:2511.08567,

    arXiv

  33. [33]

    The construction procedure follows (Zeng et al., 2024b)

    D Interleaved Speech–Text Data Construction Following (Zeng et al., 2024b; Nguyen et al., 2025), we construct interleaved speech–text data separately from the text training corpora of COSE, e-SNLI, and OpenBookQA. The construction procedure follows (Zeng et al., 2024b). Specifically, we first convert text responses into speech using the Coqui TTS API (Ere...

    2025