Pith. sign in

REVIEW 5 major objections 5 minor 70 references

PCS replaces the fixed coefficient in activation steering with a similarity-conditioned probabilistic one and reports higher direction accuracy on all ten datasets tested.

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 Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 13:52 UTC pith:NQKGKO7X

load-bearing objection PCS offers a plausible calibration idea, but the reported gains are not trustworthy: λ is tuned per prompt on the same logit-difference metric used to score accuracy. the 5 major comments →

arxiv 2607.18259 v1 pith:NQKGKO7X submitted 2026-05-15 cs.AI

Probabilistic Concept-Aware Steering for Trustworthy LLM Inference

Brian Becker , Rui Chu , Yingjie Lao This is my paper
classification cs.AI
keywords steering vectorsactivation engineeringinference-time interventionprobabilistic strength calibrationcosine similarityconcept-aware steeringlogit differencesemantic alignment
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.

This paper sets out to show that the biggest practical weakness of steering vectors—directions in a model's internal activation space that push output toward a concept—can be fixed by treating the push strength as a random sample from a Gaussian whose mean and spread are set by how similar the prompt's concept direction is to the steering direction. If that works, LLM behavior could be controlled at inference time without retuning or parameter updates, which would make fine-grained semantic control cheaper and more trustworthy. The authors report that this method, PCS, beats fixed-strength baselines on every dataset considered, by 24.8 and 16.8 percentage points on average, while cutting variance roughly in half and running with near-zero setup cost. The authors also concede that about a quarter of prompts remain resistant and that PCS currently needs a concept vector for every prompt (Appendix B.7), so the win is real but bounded.

Core claim

On its own terms, the paper advances one claim: the optimal steering strength for an LLM intervention is not a constant but a sample from a Gaussian distribution whose parameters interpolate between the orthogonality and full-alignment cases according to cosine similarity s between the steering vector and the prompt concept vector, (mu,sigma) = (1-s)(mu0,sigma0) + s(mus,sigmas), with extra variance when s falls below tau=0.2 and hard clipping on both parameters. The injected intervention is affine, h' = [I lambda v_sv] h~, applied at a chosen mid-layer. Empirically, PCS reports higher direction accuracy than fixed-strength steering baselines on all ten datasets, average gains of +24.8 and +1

What carries the argument

The central mechanism is probabilistic strength calibration: lambda, the scalar multiplying the steering vector, is drawn from N(mu,sigma^2) where both parameters are conditioned on cosine similarity between the prompt's concept vector and the steering vector. This same similarity also decides between a confident low-variance schedule and an exploratory high-variance schedule at threshold tau=0.2, with the sampled lambda then applied through an affine hidden-state transformation h'=[I|lambda v_sv]h~ at one mid-transformer layer. The machinery matters because it turns a global hyperparameter (steering strength) into a per-prompt, geometry-aware quantity, and the paper's empirical case is that

Load-bearing premise

The framework assumes cosine similarity between the steering vector and a prompt's concept vector predicts how much steering force that prompt needs; if that mapping is unreliable, the adaptive distribution loses its justification.

What would settle it

Run PCS while replacing s with shuffled or constant similarities (the paper's own constant-similarity ablation gives a 6.1-point drop); if a broader set of datasets shows no drop or a drop that vanishes, the similarity-conditioned calibration is not doing the work. Also test whether the reported reverse-U peak at moderate similarity reproduces on held-out or out-of-distribution prompts; a monotone relationship would contradict the proposed sweet-spot mechanism.

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

If this is right

  • A single cosine-similarity measurement could stand in for per-task grid search, since the Gaussian parameters are fixed across concept domains (t=3.0, d=0.5, tau=0.2).
  • Interventions become dramatically cheaper: no offline optimization phase per concept (153.4 ms setup versus 600+ seconds for a parameter-updating baseline) and stable runtime.
  • Small models benefit most; PCS steering scores exceed supervised fine-tuning on 1B and 3B parameter models, which matters for resource-constrained deployment.
  • Practitioners should avoid very large lambda: the paper finds extreme scaling creates semantic drift and incoherent outputs even when logit-difference metrics look better.
  • A residual set of prompts (roughly 25%) remains anti-steerable under any reasonable strength, so steering-vector control cannot be the only safety mechanism.

Where Pith is reading between the lines

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

  • Because the paper's own top-performing pair can have negative cosine similarity (SA-Gen with Corr-More at -0.42), raw cosine may not be the best calibration signal; a learned or ensemble estimate of expected steering efficacy could improve on the interpolation equation.
  • The reverse-U between inter-dataset similarity and direction accuracy suggests that steering success is a function of semantic distance, so a practical system could precompute sweet-spot strengths per source-target pair instead of sampling at inference.
  • PCS's dependence on a per-prompt concept vector is the main barrier to open-ended use; plugging in zero-shot concept detection or using the model's own intermediate representation as the concept vector is a natural extension the authors flag as future work.
  • The variance reduction from roughly 48-50 to 28.6 hints that the method is not just shifting means but stabilizing behavior; testing whether this stability persists on out-of-distribution prompts would clarify the generalization claim.

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

5 major / 5 minor

Summary. The paper introduces Probabilistic Concept-Aware Steering (PCS), an inference-time steering method for LLMs. For each prompt, PCS computes the cosine similarity between a concept vector and a steering vector, maps that similarity to the mean and variance of a Gaussian over the steering coefficient λ, samples k candidate λ values, evaluates each by its induced change in logit difference, and applies the best one through an affine hidden-state intervention. The authors claim large gains in direction accuracy over fixed-strength baselines (CAA and a discrete grid Λ1), improved steering scores, reduced variance, and lower inference cost, and they report experiments across multiple model families and ten datasets.

Significance. If the central mechanism were cleanly isolated, PCS would be a useful contribution: it is training-free, model-agnostic, addresses the real limitation of fixed-strength steering, and the paper provides code and uses public datasets. However, the headline evaluation grants PCS per-prompt access to the same logit-difference signal on which it is scored, while the baselines use fixed strengths. The reported margins therefore do not establish that cosine-similarity-conditioned calibration is the cause of the improvement. The inconsistent numbers between the main table and the appendix further undermine the quantitative claim. The idea is defensible and the evaluation flaw is fixable, but as written the central empirical claim is not supported.

major comments (5)
  1. [§3.4, §4, Appendix B.2.1/B.2.5] The main comparison in Table 1c is circular. Section 3.4 states that PCS 'sample[s] k candidate values from the adaptive Gaussian distribution and evaluate[s] each by measuring the resulting change in logit difference.' Direction accuracy is then defined by whether the logit difference shifts in the intended direction (Appendix B.2.1, B.2.5, threshold 0.2). Thus PCS receives online access to the exact metric on which it is scored, while CAA and Λ1 use fixed strengths. The reported +24.8/+16.8 percentage-point margins conflate per-prompt oracle search over λ with the contribution of similarity-conditioned sampling. A fair comparison must include a baseline that also optimizes λ per prompt over the same candidate set, or must select λ on training prompts and evaluate on held-out prompts.
  2. [§4, Evaluation Metrics] Accuracy is reported on a 'steerable' subset defined by significant logit shifts (s ≥ 0.2). Because PCS selects λ specifically to maximize the logit-difference shift, membership in this subset is itself influenced by the optimization objective. Filtering on the outcome variable after optimizing the same outcome can inflate measured accuracy. Please report metrics on the full test set, and if the subset analysis is retained, define steerability by an independent criterion or by an out-of-sample λ.
  3. [Table 1c vs. Table 18] The main and appendix tables contradict each other. For example, SA-Gen is 72.6% in Table 1c but 68.2% in Table 18; Coord-Ver is 72.6% vs. 56.7%; Corr-Less is 69.2% vs. 77.6%; Corr-More is 77.1% vs. 76.6%. Since Table 18 is presumably the supporting evidence for Table 1c, the quantitative central claim is not reliably established. The authors must reconcile these numbers and state which are final.
  4. [§3.4 vs. Algorithm 1] The mathematical description is internally inconsistent. Equation (2) defines (µ, σ) = (1−s)(µ0, σ0) + s(µs, σs) and mentions a κ downscaling, but Algorithm 1 uses µ = t + sα and σ = d·(w + (1−s)/v) or d + (1−s)·a, with no κ and no (µ0, σ0), (µs, σs). These quantities are never defined. This ambiguity blocks reproduction and must be fixed by unifying the notation and defining every symbol.
  5. [§3.5, Appendix B.1] The threshold τ = 0.2 is imported from SEMSCORE, where it reflects correlation with human quality judgments for textual similarity. It is then applied to a different quantity—logit-difference steerability—without evidence that the two thresholds coincide. Since τ changes the sampling schedule, the paper should provide a calibration curve or ablation over τ on steering outcomes, rather than relying on an analogy to an unrelated metric.
minor comments (5)
  1. [Section 1] The Introduction contains a duplicated paragraph beginning 'As large language models (LLMs) become increasingly prevalent...' which repeats several sentences verbatim.
  2. [Section 2/3] There are typos such as 'address address' and 'Our figue in the experiment section...' Please proofread.
  3. [Table 1a] The SFT row is malformed ('SFT –88.885.8 50.6') and does not align with the model columns.
  4. [§4.3.1] The sentence 'It can be proven that SA-NN achieves 77.1% accuracy...' is not a proof; it is an empirical observation. Please rephrase.
  5. [Table 9] The top direction-accuracy pairings include negative cosine similarities (e.g., SA-Gen with Corr-More at −0.42). This appears to conflict with the claim that cosine similarity is a reliable guide to intervention strength; please discuss this explicitly, for example by reporting the λ distribution on negative-similarity pairs.

Circularity Check

2 steps flagged

Headline PCS direction-accuracy gains are not isolated from test-time λ selection on the same logit-difference metric used for scoring.

specific steps
  1. fitted input called prediction [Section 3.4 (Eq. 2); Section 4.1; Appendix B.2.1/B.2.5]
    "To determine the optimal steering strength, we sample k candidate values from the adaptive Gaussian distribution and evaluate each by measuring the resulting change in logit difference. ... As shown in Table 1c, PCS yields higher accuracy than the two baseline steering methods, CAA and Λ1, on every dataset considered."

    Direction accuracy and effectiveness are scored by whether the logit difference (Eq. 1) moves in the intended direction (B.2.1: 'a logit difference change in the correct direction'; B.2.5: 'effective' if LD changes by ≥0.2 in the intended direction). PCS chooses its per-prompt λ by measuring exactly this logit-difference change, so it receives test-time access to the objective on which it is evaluated. CAA and Λ1 use fixed strengths, so the +24.8/+16.8 pp comparison conflates oracle search over λ with the contribution of cosine-conditioned calibration; the prediction is largely forced by selecting on the scoring metric.

  2. other [Section 4, Evaluation Metrics; Appendix B.2.5]
    "Furthermore, we conduct a fine-grained analysis of Steerable Examples, defined as the subset of prompts that demonstrate a significant response to intervention, typically characterized by significant logit shifts of ¯s≥0.2."

    The 'steerable examples' filter uses the same logit-shift criterion (≥0.2 in the intended direction) that defines effectiveness in B.2.5. Restricting the evaluation to prompts that already respond to the intervention selects on the outcome; gains reported on this subset are partly guaranteed by the inclusion rule rather than by the calibration method.

full rationale

The main circularity is evaluation-design, not derivation: PCS's per-prompt steering coefficient λ is selected by measuring the resulting change in logit difference (Section 3.4), while the reported direction accuracy / effectiveness is defined by whether that same logit difference shifts in the intended direction (B.2.1, B.2.5). The comparison against fixed-strength CAA and Λ1 therefore cannot separate the proposed cosine-conditioned Gaussian calibration from a test-time oracle over λ. The second, weaker issue is the 'Steerable Examples' filter, which conditions on the logit-shift outcome before reporting accuracy. I did not find load-bearing self-citations: the cited prior work on steering vectors, propensity curves, and the linear representation hypothesis is external, and the paper does not invoke a uniqueness theorem or an author-specific ansatz to force its conclusions. The 'negative cosine similarity yet high accuracy' Table 9 and the mismatch between Table 1c and Appendix Table 18 (e.g., SA-Gen 72.6 vs 68.2; Coord-Ver 72.6 vs 56.7) are correctness/reproducibility concerns, not additional circularity. Mitigating factors are that λ is sampled from a cosine-constrained Gaussian with only k=5 candidates, and Appendix B.2.4's cosine-ablation gives the calibration idea some independent content; hence the result is partially, not wholly, forced. Overall, the central quantitative claim partially reduces to optimizing the metric at inference time, so a moderate circularity score is warranted.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

PCS introduces no new physical or conceptual entities beyond the method itself. Its free parameters are numerous (t, d, α, τ, bounds, v, w, a, k, layer L), all hand-set or ablation-selected. The axiom ledger shows that the central claim depends on the linear representation hypothesis, cosine similarity as a calibration signal, and the external availability of concept vectors—none of which is independently established in the paper.

free parameters (8)
  • target mean t = 3.0
    Center of the steering-strength distribution; hand-set and fixed across domains (Table 4).
  • target dispersion d = 0.5
    Anchor standard deviation for λ; hand-set (Table 4).
  • similarity scaling α = 2.5
    Maps cosine similarity to the mean in Algorithm 1; hand-set (Table 4).
  • similarity threshold τ = 0.2
    Boundary for switching to exploratory σ; justified by SEMSCORE but transferred across domains (§3.5, Table 4).
  • µ and σ clip bounds = µ∈[0.5,8.0], σ∈[0.3,2.5]
    Clipping limits in Algorithm 1; hand-set (Table 4).
  • similarity weightings v, w and exploration scale a = v=2, w=0.5, a=0.5
    Ablation-selected on Corrigible-More (§B.2.6, Table 7); w=0.5 chosen as a multi-objective compromise.
  • candidate count k = 5
    Number of λ candidates sampled per prompt; no sensitivity analysis reported (Table 4).
  • intervention layer L = 16 (LLaMA3-8B), 14 (LLaMA2-7B), 14 (3B), 8 (1B), 21 (Qwen-14B)
    Layer chosen per model via layer-wise ablation (§B.1.2, Table 5); a hand-selected architectural choice.
axioms (6)
  • domain assumption Linear representation hypothesis: concepts correspond to linear directions in activation space.
    §3.1 explicitly builds on this hypothesis, citing Park et al. 2023; it is not proven here.
  • domain assumption Cosine similarity between v_sv and v_con is a reliable guide to optimal steering strength λ.
    Eq. 2 and Algorithm 1 make this the load-bearing modeling premise of PCS.
  • domain assumption A concept vector v_con can be obtained for every input prompt.
    §3.5 and B.7 require a pre-determined concept vector or external detector; authors acknowledge zero-shot limitations.
  • domain assumption Maximizing logit difference LD is a valid objective for steering quality.
    §3.2 frames steering as maximizing LD without external validation connecting LD to human judgments.
  • ad hoc to paper The SEMSCORE threshold τ=0.2 transfers to logit-difference steerability.
    §3.5 and B.1 borrow a semantic-similarity threshold from a different evaluation context without direct evidence that it applies here.
  • domain assumption LLM-as-a-judge ratings reliably measure concept adherence and fluency.
    B.1.1 builds the steering score on rule-based and LLM judges; no human agreement or calibration is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 23456 in / 14711 out tokens · 124534 ms · 2026-08-02T13:52:04.779194+00:00 · methodology

0 comments
read the original abstract

Steering vectors (SVs), an inference-time intervention technique for large language models (LLMs), guide the generation process by adding a concept-specific direction vector to intermediate activations during inference. However, existing SV methods frequently yield representation-incoherent behaviors that undermine interpretability and fine-grained control, largely because prior work has focused on binary positive-negative steering evaluation while employing discrete clustering metrics that fail to capture the continuous spectrum of semantic alignment. In this work, we present the Probabilistic Concept-Aware Steering (PCS) framework for LLM inference. PCS preserves original task competence while providing controllable, safety-oriented semantic bias through concept-driven steering-vector retrieval and probabilistic strength calibration.

Figures

Figures reproduced from arXiv: 2607.18259 by Brian Becker, Rui Chu, Yingjie Lao.

Figure 1
Figure 1. Figure 1: Illustration of the Probabilistic Concept-Aware Steering (PCS) framework. The top panel [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Direction accuracy as a function of steering [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Supplementary visualizations for our methodology. [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Data Pre-Processing for internal activations [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Layer-wise Performance remarkably consistent across concepts [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Mean entropy gap across transformer layers for LLaMA2-7B. This plot compares the [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Direction accuracy improvements across methods [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Cosine Similarities across SVs B.5 Additional Ablation Studies To complement our use of the Λ1 scaling scheme established in prior work, we conduct an ablation in which the cosine similarity between vconcept and vsteer is uniformly set to 0.8. While several datasets maintain directional gains comparable to those achieved under probabilistic scaling, we observe a consistent decline in effectiveness rates ac… view at source ↗

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

70 extracted references · 33 linked inside Pith

  1. [4]

    2026 , eprint=

    COLD-Steer: Steering Large Language Models via In-Context One-step Learning Dynamics , author=. 2026 , eprint=

    2026

  2. [5]

    2026 , eprint=

    W2S-AlignTree: Weak-to-Strong Inference-Time Alignment for Large Language Models via Monte Carlo Tree Search , author=. 2026 , eprint=

    2026

  3. [7]

    Neurocomputing , volume=

    Understanding llms: A comprehensive overview from training to inference , author=. Neurocomputing , volume=. 2025 , publisher=

    2025

  4. [8]

    Advances in Neural Information Processing Systems , volume=

    Analysing the generalisation and reliability of steering vectors , author=. Advances in Neural Information Processing Systems , volume=

  5. [10]

    arXiv preprint arXiv:2307.09288 , year=

    Llama 2: Open foundation and fine-tuned chat models , author=. arXiv preprint arXiv:2307.09288 , year=

    Pith/arXiv arXiv

  6. [12]

    arXiv preprint arXiv:2109.07958 , year=

    Truthfulqa: Measuring how models mimic human falsehoods , author=. arXiv preprint arXiv:2109.07958 , year=

    Pith/arXiv arXiv

  7. [13]

    Findings of the association for computational linguistics: ACL 2023 , pages=

    Discovering language model behaviors with model-written evaluations , author=. Findings of the association for computational linguistics: ACL 2023 , pages=

    2023

  8. [14]

    Advances in Neural Information Processing Systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , volume=

  9. [22]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  10. [23]

    arXiv preprint arXiv:2403.05767 , year=

    Extending activation steering to broad skills and multiple behaviours , author=. arXiv preprint arXiv:2403.05767 , year=

    Pith/arXiv arXiv

  11. [24]

    The 2023 Conference on Empirical Methods in Natural Language Processing , year=

    Conceptor-Aided Debiasing of Large Language Models , author=. The 2023 Conference on Empirical Methods in Natural Language Processing , year=

    2023

  12. [25]

    ICLR 2025 Workshop on Building Trust in Language Models and Applications , year=

    Understanding (Un)Reliability of Steering Vectors in Language Models , author=. ICLR 2025 Workshop on Building Trust in Language Models and Applications , year=

    2025

  13. [27]

    arXiv preprint arXiv:2101.00190 , year=

    Prefix-tuning: Optimizing continuous prompts for generation , author=. arXiv preprint arXiv:2101.00190 , year=

    Pith/arXiv arXiv

  14. [28]

    arXiv preprint arXiv:2104.08691 , year=

    The power of scale for parameter-efficient prompt tuning , author=. arXiv preprint arXiv:2104.08691 , year=

    Pith/arXiv arXiv

  15. [30]

    arXiv preprint arXiv:2311.06668 , year=

    In-context vectors: Making in context learning more effective and controllable through latent space steering , author=. arXiv preprint arXiv:2311.06668 , year=

    Pith/arXiv arXiv

  16. [31]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

  17. [32]

    Advances in Neural Information Processing Systems , volume=

    Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization , author=. Advances in Neural Information Processing Systems , volume=

  18. [36]

    Proceedings of the ACM on Web Conference 2025 , pages=

    Adaptive activation steering: A tuning-free llm truthfulness improvement method for diverse hallucinations categories , author=. Proceedings of the ACM on Web Conference 2025 , pages=

    2025

  19. [38]

    arXiv preprint arXiv:2205.05124 , year=

    Extracting latent steering vectors from pretrained language models , author=. arXiv preprint arXiv:2205.05124 , year=

    Pith/arXiv arXiv

  20. [39]

    arXiv preprint arXiv:2310.01405 , year=

    Representation engineering: A top-down approach to ai transparency , author=. arXiv preprint arXiv:2310.01405 , year=

    Pith/arXiv arXiv

  21. [40]

    arXiv preprint arXiv:2407.11606 , year=

    The foundations of tokenization: Statistical and computational concerns , author=. arXiv preprint arXiv:2407.11606 , year=

    Pith/arXiv arXiv

  22. [43]

    European conference on computer vision , pages=

    Gpt4roi: Instruction tuning large language model on region-of-interest , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  23. [46]

    Discovering Language Model Behaviors with Model-Written Evaluations , publisher =

    Perez, Ethan and Ringer, Sam and Lukošiūtė, Kamilė and Nguyen, Karina and Chen, Edwin and Heiner, Scott and Pettit, Craig and Olsson, Catherine and Kundu, Sandipan and Kadavath, Saurav and Jones, Andy and Chen, Anna and Mann, Ben and Israel, Brian and Seethor, Bryan and McKinnon, Cameron and Olah, Christopher and Yan, Da and Amodei, Daniela and Amodei, Da...

    Pith/arXiv arXiv doi:10.48550/arxiv.2212.09251 2022

  24. [47]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  25. [49]

    Neural computation , volume=

    A fast learning algorithm for deep belief nets , author=. Neural computation , volume=. 2006 , publisher=

    2006

  26. [50]

    arXiv preprint arXiv:2309.16609 , year=

    Qwen technical report , author=. arXiv preprint arXiv:2309.16609 , year=

    Pith/arXiv arXiv

  27. [51]

    Advances in Neural Information Processing Systems , volume=

    Controlmllm: Training-free visual prompt learning for multimodal large language models , author=. Advances in Neural Information Processing Systems , volume=

  28. [52]

    Advances in Neural Information Processing Systems , volume=

    Adaptable logical control for large language models , author=. Advances in Neural Information Processing Systems , volume=

  29. [53]

    Advances in Neural Information Processing Systems , volume=

    Toward efficient inference for mixture of experts , author=. Advances in Neural Information Processing Systems , volume=

  30. [54]

    Advances in neural information processing systems , volume=

    Efficient inference in fully connected crfs with gaussian edge potentials , author=. Advances in neural information processing systems , volume=

  31. [55]

    Semscore: Automated evaluation of instruction-tuned llms based on semantic textual similarity

    Ansar Aynetdinov and Alan Akbik. Semscore: Automated evaluation of instruction-tuned llms based on semantic textual similarity. arXiv preprint arXiv:2401.17072, 2024

    Pith/arXiv arXiv 2024

  32. [56]

    Explicit vulnerability generation with llms: An investigation beyond adversarial attacks

    Emir Bosnak, Sahand Moslemi, Mayasah Lami, and Anil Koyuncu. Explicit vulnerability generation with llms: An investigation beyond adversarial attacks. arXiv preprint arXiv:2507.10054, 2025

    Pith/arXiv arXiv 2025

  33. [57]

    Current and future use of large language models for knowledge work

    Michelle Brachman, Amina El-Ashry, Casey Dugan, and Werner Geyer. Current and future use of large language models for knowledge work. arXiv preprint arXiv:2503.16774, 2025

    Pith/arXiv arXiv 2025

  34. [58]

    Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization

    Yuanpu Cao, Tianrong Zhang, Bochuan Cao, Ziyi Yin, Lu Lin, Fenglong Ma, and Jinghui Chen. Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization. Advances in Neural Information Processing Systems, 37: 0 49519--49551, 2024

    2024

  35. [59]

    Enhancing fast feed forward networks with load balancing and a master leaf node

    Andreas Charalampopoulos, Nikolas Chatzis, Foivos Ntoulas-Panagiotopoulos, Charilaos Papaioannou, and Alexandros Potamianos. Enhancing fast feed forward networks with load balancing and a master leaf node. arXiv preprint arXiv:2405.16836, 2024

    Pith/arXiv arXiv 2024

  36. [60]

    Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning

    Zhongzhi Chen, Xingwu Sun, Xianfeng Jiao, Fengzong Lian, Zhanhui Kang, Di Wang, and Chengzhong Xu. Truth forest: Toward multi-scale truthfulness in large language models through intervention without tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 20967--20974, 2024

    2024

  37. [61]

    Steering off course: Reliability challenges in steering language models

    Patrick Queiroz Da Silva, Hari Sethuraman, Dheeraj Rajagopal, Hannaneh Hajishirzi, and Sachin Kumar. Steering off course: Reliability challenges in steering language models. arXiv preprint arXiv:2504.04635, 2025

    Pith/arXiv arXiv 2025

  38. [62]

    W2s-aligntree: Weak-to-strong inference-time alignment for large language models via monte carlo tree search, 2026

    Zhenyu Ding, Yuhao Wang, Tengyue Xiao, Haoying Wang, Caigui Jiang, and Ning Ding. W2s-aligntree: Weak-to-strong inference-time alignment for large language models via monte carlo tree search, 2026. URL https://arxiv.org/abs/2511.11518

    arXiv 2026

  39. [63]

    How abilities in large language models are affected by supervised fine-tuning data composition

    Guanting Dong, Hongyi Yuan, Keming Lu, Chengpeng Li, Mingfeng Xue, Dayiheng Liu, Wei Wang, Zheng Yuan, Chang Zhou, and Jingren Zhou. How abilities in large language models are affected by supervised fine-tuning data composition. arXiv preprint arXiv:2310.05492, 2023

    Pith/arXiv arXiv 2023

  40. [64]

    Differentially private steering for large language model alignment

    Anmol Goel, Yaxi Hu, Iryna Gurevych, and Amartya Sanyal. Differentially private steering for large language model alignment. arXiv preprint arXiv:2501.18532, 2025

    Pith/arXiv arXiv 2025

  41. [65]

    Safeswitch: Steering unsafe llm behavior via internal activation signals

    Peixuan Han, Cheng Qian, Xiusi Chen, Yuji Zhang, Denghui Zhang, and Heng Ji. Safeswitch: Steering unsafe llm behavior via internal activation signals. arXiv preprint arXiv:2502.01042, 2025

    arXiv 2025

  42. [66]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1 0 (2): 0 3, 2022

    2022

  43. [67]

    Toward efficient inference for mixture of experts

    Haiyang Huang, Newsha Ardalani, Anna Sun, Liu Ke, Shruti Bhosale, Hsien-Hsin Lee, Carole-Jean Wu, and Benjamin Lee. Toward efficient inference for mixture of experts. Advances in Neural Information Processing Systems, 37: 0 84033--84059, 2024

    2024

  44. [68]

    Enhancing instruction following of llms via activation steering with dynamic rejection

    Minjae Kang and Jaehyung Kim. Enhancing instruction following of llms via activation steering with dynamic rejection. arXiv preprint arXiv:2603.06745, 2026

    arXiv 2026

  45. [69]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  46. [70]

    Style vectors for steering generative large language model

    Kai Konen, Sophie Jentzsch, Diaoul \'e Diallo, Peer Sch \"u tt, Oliver Bensch, Roxanne El Baff, Dominik Opitz, and Tobias Hecking. Style vectors for steering generative large language model. arXiv preprint arXiv:2402.01618, 2024

    Pith/arXiv arXiv 2024

  47. [71]

    a henb \

    Philipp Kr \"a henb \"u hl and Vladlen Koltun. Efficient inference in fully connected crfs with gaussian edge potentials. Advances in neural information processing systems, 24, 2011

    2011

  48. [72]

    Contextualization distillation from large language model for knowledge graph completion

    Dawei Li, Zhen Tan, Tianlong Chen, and Huan Liu. Contextualization distillation from large language model for knowledge graph completion. arXiv preprint arXiv:2402.01729, 2024

    Pith/arXiv arXiv 2024

  49. [73]

    Inference-time intervention: Eliciting truthful answers from a language model

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36: 0 41451--41530, 2023

    2023

  50. [74]

    Understanding llms: A comprehensive overview from training to inference

    Yiheng Liu, Hao He, Tianle Han, Xu Zhang, Mengyuan Liu, Jiaming Tian, Yutong Zhang, Jiaqi Wang, Xiaohui Gao, Tianyang Zhong, et al. Understanding llms: A comprehensive overview from training to inference. Neurocomputing, 620: 0 129190, 2025

    2025

  51. [75]

    Steering llama 2 via contrastive activation addition

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681, 2023

    Pith/arXiv arXiv 2023

  52. [76]

    Unveiling the secret recipe: A guide for supervised fine-tuning small llms

    Aldo Pareja, Nikhil Shivakumar Nayak, Hao Wang, Krishnateja Killamsetty, Shivchander Sudalairaj, Wenlong Zhao, Seungwook Han, Abhishek Bhandwaldar, Guangxuan Xu, Kai Xu, et al. Unveiling the secret recipe: A guide for supervised fine-tuning small llms. arXiv preprint arXiv:2412.13337, 2024

    Pith/arXiv arXiv 2024

  53. [77]

    The linear representation hypothesis and the geometry of large language models

    Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. arXiv preprint arXiv:2311.03658, 2023

    Pith/arXiv arXiv 2023

  54. [78]

    Discovering language model behaviors with model-written evaluations

    Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kadavath, et al. Discovering language model behaviors with model-written evaluations. In Findings of the association for computational linguistics: ACL 2023, pages 13387--13434, 2023

    2023

  55. [79]

    Steering large language models using conceptors: Improving addition-based activation engineering

    Joris Postmus and Steven Abreu. Steering large language models using conceptors: Improving addition-based activation engineering. arXiv preprint arXiv:2410.16314, 2024

    Pith/arXiv arXiv 2024

  56. [80]

    Towards reliable evaluation of behavior steering interventions in llms

    Itamar Pres, Laura Ruis, Ekdeep Singh Lubana, and David Krueger. Towards reliable evaluation of behavior steering interventions in llms. arXiv preprint arXiv:2410.17245, 2024

    Pith/arXiv arXiv 2024

  57. [81]

    Kartik Sharma and Rakshit S. Trivedi. Cold-steer: Steering large language models via in-context one-step learning dynamics, 2026. URL https://arxiv.org/abs/2603.06495

    arXiv 2026

  58. [82]

    Enstom: Enhancing dialogue systems with entropy-scaled steering vectors for topic maintenance

    Heejae Suh, Yejin Jeon, Deokhyung Kang, Taehee Park, Yejin Min, and Gary Geunbae Lee. Enstom: Enhancing dialogue systems with entropy-scaled steering vectors for topic maintenance. arXiv preprint arXiv:2505.16526, 2025

    Pith/arXiv arXiv 2025

  59. [83]

    Analysing the generalisation and reliability of steering vectors

    Daniel Tan, David Chanin, Aengus Lynch, Brooks Paige, Dimitrios Kanoulas, Adri \`a Garriga-Alonso, and Robert Kirk. Analysing the generalisation and reliability of steering vectors. Advances in Neural Information Processing Systems, 37: 0 139179--139212, 2024

    2024

  60. [84]

    Steering language models with activation engineering

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. Steering language models with activation engineering. arXiv preprint arXiv:2308.10248, 2023

    Pith/arXiv arXiv 2023

  61. [85]

    Evolving prompts in-context: An open-ended, self-replicating perspective

    Jianyu Wang, Zhiqiang Hu, and Lidong Bing. Evolving prompts in-context: An open-ended, self-replicating perspective. arXiv preprint arXiv:2506.17930, 2025 a

    Pith/arXiv arXiv 2025

  62. [86]

    Adaptive activation steering: A tuning-free llm truthfulness improvement method for diverse hallucinations categories

    Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma. Adaptive activation steering: A tuning-free llm truthfulness improvement method for diverse hallucinations categories. In Proceedings of the ACM on Web Conference 2025, pages 2562--2578, 2025 b

    2025

  63. [87]

    Semantics-adaptive activation intervention for llms via dynamic steering vectors

    Weixuan Wang, Jingyuan Yang, and Wei Peng. Semantics-adaptive activation intervention for llms via dynamic steering vectors. arXiv preprint arXiv:2410.12299, 2024

    Pith/arXiv arXiv 2024

  64. [88]

    Controlmllm: Training-free visual prompt learning for multimodal large language models

    Mingrui Wu, Xinyue Cai, Jiayi Ji, Jiale Li, Oucheng Huang, Gen Luo, Hao Fei, Guannan Jiang, Xiaoshuai Sun, and Rongrong Ji. Controlmllm: Training-free visual prompt learning for multimodal large language models. Advances in Neural Information Processing Systems, 37: 0 45206--45234, 2024

    2024

  65. [89]

    Axbench: Steering llms? even simple baselines outperform sparse autoencoders

    Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D Manning, and Christopher Potts. Axbench: Steering llms? even simple baselines outperform sparse autoencoders. arXiv preprint arXiv:2501.17148, 2025 a

    Pith/arXiv arXiv 2025

  66. [90]

    Improved representation steering for language models

    Zhengxuan Wu, Qinan Yu, Aryaman Arora, Christopher D Manning, and Christopher Potts. Improved representation steering for language models. arXiv preprint arXiv:2505.20809, 2025 b

    Pith/arXiv arXiv 2025

  67. [91]

    Yifei, Lyle Ungar, and Jo \ a o Sedoc

    Li S. Yifei, Lyle Ungar, and Jo \ a o Sedoc. Conceptor-aided debiasing of large language models. In The 2023 Conference on Empirical Methods in Natural Language Processing, 2023. URL https://openreview.net/forum?id=M6BJfQ9oup

    2023

  68. [92]

    Adaptable logical control for large language models

    Honghua Zhang, Po-Nien Kung, Masahiro Yoshida, Guy Van den Broeck, and Nanyun Peng. Adaptable logical control for large language models. Advances in Neural Information Processing Systems, 37: 0 115563--115587, 2024 a

    2024

  69. [93]

    Gpt4roi: Instruction tuning large language model on region-of-interest

    Shilong Zhang, Peize Sun, Shoufa Chen, Min Xiao, Wenqi Shao, Wenwei Zhang, Yu Liu, Kai Chen, and Ping Luo. Gpt4roi: Instruction tuning large language model on region-of-interest. In European conference on computer vision, pages 52--70. Springer, 2024 b

    2024

  70. [94]

    Batch calibration: Rethinking calibration for in-context learning and prompt engineering

    Han Zhou, Xingchen Wan, Lev Proleev, Diana Mincu, Jilin Chen, Katherine Heller, and Subhrajit Roy. Batch calibration: Rethinking calibration for in-context learning and prompt engineering. arXiv preprint arXiv:2309.17249, 2023

    Pith/arXiv arXiv 2023