Pith. sign in

REVIEW 3 major objections 6 minor 5 cited by

GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Gradient-based attribution identifies the tokens responsible for bad outputs and steers models away from them, outperforming fine-tuning on truthfulness, toxicity, and hallucination benchmarks while preserving general capability.

desk verdict A solid integration of IG-based attribution and contrastive steering that overclaims slightly and uses a partly circular VLM metric, but deserves a serious referee. read the letter →

arxiv 2507.18043 v2 pith:BCQSLZDR submitted 2025-07-24 cs.CL cs.AIcs.CV

Duy Nguyen , Archiki Prasad , Elias Stengel-Eskin , Mohit Bansal This is my paper
classification cs.CLcs.AIcs.CV
keywords inference-timesteeringactivationIntegratedGradientstokenattributionvision-languagemodelshallucinationmitigationtruthfulnessmodelalignment
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

GRAINS is an inference-time steering method that tries to establish that token-level attribution can replace global, fixed steering vectors. It uses Integrated Gradients on a contrastive preference objective—the log-likelihood difference between a preferred and a dispreferred response—to find the tokens most responsible for each behavior, then builds layer-wise steering vectors from the activation shifts caused by ablating those tokens. The paper reports that this selective intervention beats both LoRA fine-tuning and existing steering baselines: +13.22% accuracy on TruthfulQA with Llama-3.1-8B, a hallucination-rate drop from 0.624 to 0.514 on MMHal-Bench with LLaVA-1.6-7B, and +8.11% alignment win rate on SPA-VL, while leaving MMLU and MMMU nearly unchanged. If true, it would make alignment a lightweight, interpretable, per-task edit rather than a retraining procedure, and it would connect attribution research directly to model control.

What carries the argument

Contrastive Integrated Gradients form the attribution step: for each input token, IG integrates gradients of the preference loss $f(x)$ along a path from a baseline to the input, yielding a signed score; positive scores mark tokens driving preferred outputs and negative scores mark tokens driving dispreferred outputs. The top-k sets $I^+_k(x)$ and $I^-_k(x)$ are ablated to produce contrastive inputs $x\setminus I^+$ and $x\setminus I^-$, whose last-token hidden states give delta vectors $\delta^+_l = h^{(l)}_{\text{last}}(x) - h^{(l)}_{\text{last}}(x\setminus I^+)$ and $\delta^-_l = h^{(l)}_{\text{last}}(x) - h^{(l)}_{\text{last}}(x\setminus I^-)$. PCA over a 50-sample set extracts $v_l^+$ and $v_l^-$, the steering vector is $v_l = v_l^+ - v_l^-$, and at inference it is added to every hidden state as $\tilde{h}_{t,l} = h_{t,l} + \lambda v_l$, then rescaled to the original norm via Eq. (8).

What would settle it

Shuffle the token-attribution scores across the 50 steering samples before selecting the top-k tokens, keeping the rest of the pipeline identical; if the shuffled version matches GRAINS on TruthfulQA and MMHal-Bench within noise, the attribution step is not what causes the improvement and the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the causally influential tokens—identified by contrastive Integrated Gradients as the top-k positive and top-k negative contributors to the preference score $f(x) = \log P_\theta(y_{\text{pos}}|x) - \log P_\theta(y_{\text{neg}}|x)$—carry enough information to construct effective steering vectors. Ablating these tokens and taking the difference in last-token hidden states gives per-example delta vectors; PCA aggregates them into $v_l^+$ and $v_l^-$, and the final steering direction is $v_l = v_l^+ - v_l^-$, applied additively with norm normalization during inference. The paper argues that this vector is more targeted than a global activation shift, and that this targeting is why GRAINS outperforms fine-tuning and baselines on truthfulness, toxicity, context faithfulness, hallucination, and safety preference while preserving fluency and general reasoning.

Load-bearing premise

A single PCA direction computed from the contrastive activation shifts of 50 samples, applied uniformly to every token at every layer, transfers to held-out inputs and shifts behavior in the intended direction without damaging other capabilities.

Editorial extensions

If this is right

  • Steering becomes a per-objective, per-layer additive edit: the same 50-sample construction routine transfers across Llama-3.1-8B, Qwen2.5-7B, LLaVA-1.6-7B, Qwen2.5-VL-7B, and Gemma-3-12B without retraining.
  • Because the intervention is localized to attributed tokens, general capabilities such as MMLU and MMMU accuracy remain near baseline, whereas global vector methods such as CAA can drop MMLU by 17.78 percentage points.
  • The contrastive preference objective lets the same machinery address truthfulness, toxicity, context faithfulness, hallucination, and safety preference, and a single-reference variant still works when preference pairs are unavailable.
  • Attribution quality is load-bearing: Integrated Gradients outperforms vanilla gradients, SmoothGrad, and random token selection by 4 to 7 points on average, so the token-selection step is what carries the gain.
  • The method adds no auxiliary modules or supervision at inference: no object detectors, no token sampling, and no extra forward passes beyond a one-time vector construction.

Reading between the lines

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

  • The binary top-k selection could be replaced by continuous attribution-weighted interpolation of hidden states, which may improve performance on inputs where influence is diffuse rather than concentrated in a few tokens; the paper does not test this variant.
  • If the PCA direction $v_l^+ - v_l^-$ is a genuine behavioral axis, then multiple such vectors built for different attributes (truthfulness, safety, style) should compose additively; the paper's vector formulation makes multi-attribute composition a direct next experiment, though it is not tested here.
  • The method's reliance on preference pairs suggests a natural extension: generate contrastive pairs from the model's own decoding candidates (for example, a hallucinated and a grounded answer) and steer without any human-labeled preference data.
  • Because construction uses only 50 samples and applies the same vector at every layer, the approach implicitly assumes the direction is model-wide rather than context-specific; measuring per-layer cosine similarity across tasks would show whether layer-wise tuning could further improve transfer.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes GrAInS, an inference-time steering method for LLMs and VLMs. The method first uses contrastive Integrated Gradients over a preference-based loss f(x) = log P(y_pos|x) - log P(y_neg|x) to identify the top-k positively and negatively attributed input tokens (Eqs. 1-3). It then constructs contrastive inputs by masking these tokens, extracts layer-wise hidden-state differences, and applies PCA to obtain steering vectors v_l = v+_l - v-_l (Eqs. 4-6). At inference, the vectors are added to hidden states with norm normalization (Eqs. 7-8). The authors evaluate on TruthfulQA, Toxigen, and FaithEval for LLMs, and on MMHal-Bench and SPA-VL for VLMs, also reporting MMLU/MMMU and BLEU-based generation-quality checks. They claim consistent improvements over LoRA fine-tuning and steering baselines such as CAA, ICV, NL-ITI, VTI, and ICT.

Significance. If the empirical claims are sustained, GrAInS would be a practically valuable contribution: it offers an interpretable, token-selective steering mechanism that does not require retraining or external modules, with modest reported construction runtime and a public code release. The ablations comparing IG against vanilla gradients, SmoothGrad, random selection, and modality-specific variants are informative and lend support to the core design choices. However, the paper's headline claim of consistent superiority is not currently supported by its own tables, and the primary multimodal alignment metric is partially circular with the training objective. The method is plausible, but the evidence as presented needs correction and strengthening.

major comments (3)
  1. [Table 2 / §4.2 / Abstract] The abstract states that GrAInS 'consistently outperforms both fine-tuning and existing steering baselines,' but Table 2 contradicts this: on Qwen2.5-VL-7B with MMHal-Bench, LoRA achieves a hallucination rate of 0.461, which is lower (better) than GrAInS's 0.473. The accompanying sentence that 'GRAINS achieves the lowest hallucination rates across all three VLMs on MMHal-Bench' is not supported by the printed numbers. This is a load-bearing inconsistency because the universal claim rests on this table. The authors must either correct a typo in the table or substantially weaken the claim; in addition, no error bars or multiple seeds are reported, so even the remaining margins (e.g., 0.473 vs. 0.499 for VTI) cannot be assessed for statistical significance.
  2. [§3.1 / §4.2 / Table 3] The SPA-VL evaluation metric is the percentage of cases where the chosen response receives higher log-likelihood than the rejected response, which is exactly the sign of the contrastive objective f(x) defined in Eq. (1). The steering vectors are constructed in Eqs. (4)-(6) to shift activations in the direction that increases this objective, so the SPA-VL 'win rates' in Table 3 partly measure the method's own optimization target rather than an independent improvement in generation quality. I recommend reporting a true generation win rate, such as preference judgments over sampled outputs, or an independent held-out metric, to establish that the alignment gains are not circular.
  3. [§4.1 / §4.2 / §A.2] All headline results are single runs with no error bars, confidence intervals, or significance tests. The hyperparameters k and alpha are tuned on dev splits of the same benchmarks (Figures 5-6), and Table 1/2/3 report one value per setting. Given that several claimed improvements are small (e.g., FaithEval on Llama: 70.94 vs. LoRA's 69.93; MMHal on Qwen: 0.473 vs. VTI's 0.499), the authors should report multiple seeds or bootstrap intervals before claiming consistent superiority over baselines.
minor comments (6)
  1. [Fig. 1 caption] The phrase 'constructs constructs' contains a duplicated word; please revise.
  2. [§4.1] The model name 'Qwen2.5-7B-Intruct' appears to be a typo for 'Qwen2.5-7B-Instruct'.
  3. [Eqs. (2)-(3)] The notation x_j is used both for a token embedding in Eq. (2) and for a token in Eq. (3); please clarify whether x_j denotes the token or its embedding throughout.
  4. [§3.3 / §A.2] The steering strength hyperparameter is denoted lambda in Eq. (7) but alpha in Appendix A.2 and Figures 5-6; please unify the symbol.
  5. [Table 1] Several entries are missing spacing, e.g., '59.9864.96' and '62.1264.77', making the table hard to read; please fix the formatting.
  6. [Tables 4-5] The definition of 'BLEU accuracy' is nonstandard; please specify exactly how the BLEU score is used to decide whether a generated output is closer to the positive reference than to the negative reference.

Circularity Check

1 steps flagged · score 5.0 of 10

SPA-VL win-rate is the method's own preference objective measured on held-out examples; independent MMLU/MMMU and MMHal results keep the paper from being fully circular.

  1. fitted input called prediction [Section 4.2 (SPA-VL datasets) vs Section 3.1, Eq. (1); also Eqs. (4)-(7)]
    ""Each sample includes a chosen (preferred) and rejected (dispreferred) response. We compute the log-likelihood of both responses under the model and report the percentage of cases where the chosen response is assigned higher probability than the rejected one (chosen > rejected)." ... "we define the attribution objective using a preference-based loss: f(x) = log Pθ(ypos |x) − log Pθ(yneg |x) (1)""

    The SPA-VL win rate is exactly P(f(x) > 0), with ypos/yneg the chosen/rejected responses from Eq. (1). GRAINS builds v_l (Eq. 6) from Integrated Gradients of this same f, selecting tokens that maximize f and applying v_l at inference (Eq. 7). Hence the reported 8.11% SPA-VL gain is a held-out measure of the very contrastive preference quantity the steering vector was constructed to increase. It is not an independent external alignment benchmark; it is the method's own training objective evaluated on a held-out split. The held-out split and PCA aggregation provide real generalization content, so the result is not fully forced, but the 'alignment win rate' claim partially reduces to the method's own target.

full rationale

The paper's construction is otherwise self-contained: Integrated Gradients is a standard external method, the steering vectors are obtained by PCA over contrastive activation differences, and the inference-time intervention is a simple additive shift with norm normalization. There is no load-bearing self-citation chain or imported uniqueness theorem. The main circularity concern is the SPA-VL evaluation: the paper defines its attribution objective in Eq. (1) as the same chosen-vs-rejected log-likelihood difference that SPA-VL uses to compute win rates, and the steering vector is explicitly built to increase that objective. Thus the SPA-VL result is partially a generalization check of the method's own optimization target rather than an independent alignment benchmark. Independent evidence (MMHal-Bench hallucination rates, MMLU/MMMU preservation, BLEU-accuracy generation checks) partially offsets this, so the paper is not wholly circular. Separately, the printed Table 2 shows LoRA (0.461) beating GRAINS (0.473) on Qwen2.5-VL-7B MMHal-Bench, contradicting the abstract's 'consistently outperforms both fine-tuning' claim; this is a correctness/consistency issue, not a circularity issue.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unverified modeling assumptions: that IG attributions capture causal token influence, that PCA over 50-sample contrastive deltas yields a transferable steering direction, and that adding a single global vector to all layers and positions preserves capabilities. No formal derivation or theoretical guarantee is provided; support is empirical. The method's main free parameters are k and the steering strength, which are tuned on dev splits of the evaluation benchmarks.

free parameters (3)
  • k (number of top tokens removed) = k=3 for LLMs, k=20 for VLMs
    Selected based on dev-set performance (Section A.2, Fig. 6).
  • lambda / alpha (steering strength) = Not specified for main tables; dev analysis suggests alpha=6 for Llama-3.1-8B and alpha=10 for Qwen2.5-7B on TruthfulQA
    Controls intervention magnitude in Eq. (7); tuned on dev split.
  • Number of IG integration steps = 5 for LLaVA/Qwen, 10 for Gemma
    Chosen for computational stability; not treated as a tuned hyperparameter.
assumptions (4)
  • domain assumption Token-level Integrated Gradients attributions faithfully identify the input tokens most causally responsible for the model's preference between y_pos and y_neg.
    Used in Section 3.1 to define I+ and I-; supported only by ablation (random vs IG), not by causal verification.
  • domain assumption The first principal component of the contrastive activation deltas (delta+, delta-) captures a generalizable semantic direction in residual stream that transfers across inputs.
    PCA in Eq. (5) aggregates vectors across 50 training samples; the transferability to held-out inputs is assumed.
  • domain assumption Adding a fixed steering vector to all token positions at all layers, followed by norm normalization, is a valid intervention that does not degrade model capabilities.
    Eqs. (7)-(8); no theoretical guarantee, only empirical MMLU/MMMU checks.
  • domain assumption Preference pairs (y_pos, y_neg) used to define f(x) are representative of the desired/undesired behavior for the whole evaluation set.
    Only 50 samples are used to construct vectors (Sections 4.1 and 4.2).

how reviews work

0 comments
Cite this review

Pith. "Pith review of GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs." pith.science (2026). https://pith.science/paper/BCQSLZDR

@misc{pith2026250718043,
  author       = {Pith},
  title        = {Pith review of: GrAInS: Gradient-based Attribution for Inference-Time Steering of LLMs and VLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BCQSLZDR}},
  note         = {Machine review of arXiv:2507.18043}
}
read the original abstract

Inference-time steering methods offer a lightweight alternative to fine-tuning large language models (LLMs) and vision-language models (VLMs) by modifying internal activations at test time without updating model weights. However, most existing approaches rely on fixed, global intervention vectors, overlook the causal influence of individual input tokens, and fail to leverage informative gradients from the model's logits, particularly in multimodal settings where visual and textual inputs contribute unevenly. To address these limitations, we introduce GrAInS, an inference-time steering approach that operates across both language-only and vision-language models and tasks. GrAInS uses contrastive, gradient-based attribution via Integrated Gradients to identify the top-k most influential tokens, both positively and negatively attributed based on their contribution to preferred versus dispreferred outputs. These tokens are then used to construct directional steering vectors that capture semantic shifts from undesirable to desirable behavior. During inference, GrAInS adjusts hidden activations at transformer layers guided by token-level attribution signals, and normalizes activations to preserve representational scale. This enables fine-grained, interpretable, and modular control over model behavior, without retraining or auxiliary supervision. Empirically, GrAInS consistently outperforms both fine-tuning and existing steering baselines: it achieves a 13.22% accuracy gain on TruthfulQA using Llama-3.1-8B, reduces hallucination rates on MMHal-Bench from 0.624 to 0.514 with LLaVA-1.6-7B, and improves alignment win rates on SPA-VL by 8.11%, all while preserving the model's fluency and general capabilities.

Figures

Figures reproduced from arXiv: 2507.18043 by the authors.

Figure 1
Figure 1. Comparison of prior steering methods vs. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of our attribution-guided steering method for VLMs. Our method consists of three [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Effect on preference difference ∆ = log Pθ(ypos | x) − log Pθ(yneg | x) after ablating top-k tokens based on signed Integrated Gradients. Removing tokens with high negative attribution substantially increases model preference for aligned outputs (ypos), whereas removing high positive tokens leads to preference drops. Results shown for Llama-3.1-8B-Instruct and Qwen-2.5-7B-Instruct models on TruthfulQA. f(x) = log Pθ… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Examples from MMHal-Bench demonstrating the effectiveness of [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Effect of steering strength α on model accuracy for LLaMA-3.1-8B and Qwen-2.5-7B on TruthfulQA. Larger α leads to stronger intervention; performance peaks at moderate values for Llama, while Qwen continues improving up to α = 10. 1 2 3 4 5 6 7 8 9 K 0.40 0.45 0.50 0.55…
Figure 6
Figure 6. Figure 6: Effect of token count k on model accuracy for LLaMA-3.1-8B and Qwen-2.5-7B on TruthfulQA. With a small number of important tokens, the method yields the strongest improvements. Accuracy peaks at k = 3 for LLaMA and k = 5 for Qwen before declining with larger k. model p…
Figure 7
Figure 7. Figure 7: Qualitative results of LLaVA-1.6-7B for our method and steering baselines on MMHal [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Activation Steering for Synthetic Data Generation: The Role of Diversity in Downstream Safety Detection

    cs.LG 2026-05 unverdicted novelty 6.0 of 10

    Activation steering produces synthetic safety-violating data that improves downstream classifiers over prompting on most tested concepts when a harmonic mean of alignment, coherence, and diversity is optimized.

  2. Continuous Interpretive Steering for Scalar Diversity

    cs.CL 2026-04 unverdicted novelty 6.0 of 10

    Continuous Interpretive Steering and the GraSD dataset reveal that LLMs encode graded sensitivity to scalar diversity in their internal representations, recoverable via controlled activation interventions.

  3. The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment

    cs.LG 2026-04 unverdicted novelty 6.0 of 10

    The Master Key Hypothesis states that capabilities are low-dimensional directions transferable across models through linear subspace alignment, with UNLOCK demonstrating gains such as 12.1% accuracy improvement on MAT...

  4. T2T-VICL: Cross-Task Visual In-Context Learning via Implicit Text-Driven VLMs

    cs.CV 2025-11 reject novelty 6.0 of 10

    T2T-VICL generates implicit text prompts from a mismatched demonstration pair and query to enable cross-task visual in-context learning in a frozen VLM, reporting gains over fixed prompting that are partly undermined ...

  5. Locate, Steer, and Improve: A Practical Survey of Actionable Mechanistic Interpretability in Large Language Models

    cs.CL 2026-01 unverdicted novelty 5.0 of 10

    The survey organizes mechanistic interpretability techniques into a Locate-Steer-Improve framework to enable actionable improvements in LLM alignment, capability, and efficiency.

Reference graph

Works this paper leans on

55 extracted references · 24 canonical work pages · cited by 5 Pith papers

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    A diagnostic study of explainability techniques for text classification

    Pepa Atanasova, Jakob Grue Simonsen, Christina Lioma, and Isabelle Augenstein. A diagnostic study of explainability techniques for text classification. In Bonnie Webber, Trevor Cohn, Yulan He, and Yang Liu, editors, Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 3256--3274, Online, November 2020. Asso...

    work page doi:10.18653/v1/2020.emnlp-main.263 2020

  3. [3]

    Madtp: Multimodal alignment-guided dynamic token pruning for accelerating vision-language transformer

    Jianjian Cao, Peng Ye, Shengze Li, Chong Yu, Yansong Tang, Jiwen Lu, and Tao Chen. Madtp: Multimodal alignment-guided dynamic token pruning for accelerating vision-language transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15710--15719, 2024

    work page 2024

  4. [4]

    Xprompt: Explaining large language model's generation via joint prompt attribution

    Yurui Chang, Bochuan Cao, Yujia Wang, Jinghui Chen, and Lu Lin. Xprompt: Explaining large language model's generation via joint prompt attribution. arXiv preprint arXiv:2405.20404, 2024

    work page arXiv 2024

  5. [5]

    Ict: Image-object cross-level trusted intervention for mitigating object hallucination in large vision-language models, 2024 a

    Junzhe Chen, Tianshu Zhang, Shiyu Huang, Yuwei Niu, Linfeng Zhang, Lijie Wen, and Xuming Hu. Ict: Image-object cross-level trusted intervention for mitigating object hallucination in large vision-language models, 2024 a . URL https://arxiv.org/abs/2411.15268

    arXiv 2024

  6. [6]

    Quantifying and Mitigating Unimodal Biases in Multimodal Large Language Models: A Causal Perspective

    Meiqi Chen, Yixin Cao, Yan Zhang, and Chaochao Lu. Quantifying and mitigating unimodal biases in multimodal large language models: A causal perspective. arXiv preprint arXiv:2403.18346, 2024 b

    work page Pith review arXiv 2024

  7. [7]

    Covert, Scott Lundberg, and Su-In Lee

    Ian C. Covert, Scott Lundberg, and Su-In Lee. Explaining by removing: a unified framework for model explanation. J. Mach. Learn. Res., 22 0 (1), January 2021. ISSN 1532-4435

    work page 2021

  8. [8]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    arXiv 2024

Show all 55 references
  1. [9]

    Frustratingly easy test-time adaptation of vision-language models

    Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimiliano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=e...

    2024

  2. [10]

    Immune: Improving safety against jailbreaks in multi-modal llms via inference-time alignment

    Soumya Suvra Ghosal, Souradip Chakraborty, Vaibhav Singh, Tianrui Guan, Mengdi Wang, Ahmad Beirami, Furong Huang, Alvaro Velasquez, Dinesh Manocha, and Amrit Singh Bedi. Immune: Improving safety against jailbreaks in multi-modal llms via inference-time alignment. arXiv preprin...

    2024 arXiv

  3. [11]

    T oxi G en: A large-scale machine-generated dataset for adversarial and implicit hate speech detection

    Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar. T oxi G en: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio, editors, Proceedings...

    2022 doi

  4. [12]

    Measuring massive multitask language understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representations (ICLR), 2021

    2021

  5. [13]

    Non-linear inference time intervention: Improving llm truthfulness

    Jakub Hoscilowicz, Adam Wiacek, Jan Chojnacki, Adam Cieslak, Leszek Michon, and Artur Janicki. Non-linear inference time intervention: Improving llm truthfulness. In Proc. Interspeech 2024, pages 4094--4098, 2024

    2024

  6. [14]

    Lo RA : Low-rank adaptation of large language models

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lo RA : Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=nZeVKeeFYf9

    2022

  7. [15]

    Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P....

    2024

  8. [16]

    A unified understanding and evaluation of steering methods

    Shawn Im and Yixuan Li. A unified understanding and evaluation of steering methods. arXiv preprint arXiv:2502.02716, 2025

    2025

  9. [17]

    Guided integrated gradients: An adaptive path method for removing noise

    Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, and Tolga Bolukbasi. Guided integrated gradients: An adaptive path method for removing noise. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5050--5058, 2021

    2021

  10. [18]

    Analyzing fine-tuning representation shift for multimodal llms steering alignment

    Pegah Khayatan, Mustafa Shukor, Jayneel Parekh, and Matthieu Cord. Analyzing fine-tuning representation shift for multimodal llms steering alignment. arXiv preprint arXiv:2501.03012, 2025

    2025 arXiv

  11. [19]

    Mitigating object hallucinations in large vision-language models through visual contrastive decoding

    Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...

    2024

  12. [20]

    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, 2024

    2024

  13. [21]

    Learning without forgetting

    Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE transactions on pattern analysis and machine intelligence, 40 0 (12): 0 2935--2947, 2017

    2017

  14. [22]

    T ruthful QA : Measuring how models mimic human falsehoods

    Stephanie Lin, Jacob Hilton, and Owain Evans. T ruthful QA : Measuring how models mimic human falsehoods. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio, editors, Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Lo...

    2022 doi

  15. [23]

    Zihao Lin, Samyadeep Basu, Mohammad Beigi, Varun Manjunatha, Ryan A. Rossi, Zichao Wang, Yufan Zhou, Sriram Balasubramanian, Arman Zarei, Keivan Rezaei, Ying Shen, Barry Menglong Yao, Zhiyang Xu, Qin Liu, Yuxiang Zhang, Yan Sun, Shilong Liu, Li Shen, Hongxuan Li, Soheil Feizi,...

    2025 arXiv

  16. [24]

    Llava-next: Improved reasoning, ocr, and world knowledge, January 2024 a

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-next: Improved reasoning, ocr, and world knowledge, January 2024 a . URL https://llava-vl.github.io/blog/2024-01-30-llava-next/

    2024

  17. [25]

    Reducing hallucinations in vision-language models via latent space steering

    Sheng Liu, Haotian Ye, Lei Xing, and James Zou. Reducing hallucinations in vision-language models via latent space steering. arXiv preprint arXiv:2410.15778, 2024 b

    2024 arXiv

  18. [26]

    In-context vectors: Making in context learning more effective and controllable through latent space steering

    Sheng Liu, Haotian Ye, Lei Xing, and James Y Zou. In-context vectors: Making in context learning more effective and controllable through latent space steering. In Forty-first International Conference on Machine Learning, 2024 c

    2024

  19. [27]

    Gradient episodic memory for continual learning

    David Lopez-Paz and Marc'Aurelio Ranzato. Gradient episodic memory for continual learning. Advances in neural information processing systems, 30, 2017

    2017

  20. [28]

    The geometry of truth: Emergent linear structure in large language model representations of true/false datasets

    Samuel Marks and Max Tegmark. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824, 2023

    2023 arXiv

  21. [29]

    Faitheval: Can your language model stay faithful to context, even if ''the moon is made of marshmallows''

    Yifei Ming, Senthil Purushwalkam, Shrey Pandit, Zixuan Ke, Xuan-Phi Nguyen, Caiming Xiong, and Shafiq Joty. Faitheval: Can your language model stay faithful to context, even if ''the moon is made of marshmallows''. In The Thirteenth International Conference on Learning Represe...

    2025

  22. [30]

    Risk-aware distributional intervention policies for language models

    Bao Nguyen, Binh Nguyen, Duy Nguyen, and Viet Anh Nguyen. Risk-aware distributional intervention policies for language models. arXiv preprint arXiv:2501.15758, 2025 a

    2025

  23. [31]

    Multi-attribute steering of language models via targeted intervention, 2025 b

    Duy Nguyen, Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Multi-attribute steering of language models via targeted intervention, 2025 b . URL https://arxiv.org/abs/2502.12446

    2025 arXiv

  24. [32]

    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

    2023 arXiv

  25. [33]

    Combining feature and instance attribution to detect artifacts

    Pouya Pezeshkpour, Sarthak Jain, Sameer Singh, and Byron Wallace. Combining feature and instance attribution to detect artifacts. In Smaranda Muresan, Preslav Nakov, and Aline Villavicencio, editors, Findings of the Association for Computational Linguistics: ACL 2022, pages 19...

    2022 doi

  26. [34]

    Householder pseudo-rotation: A novel approach to activation editing in LLM s with direction-magnitude perspective

    Van-Cuong Pham and Thien Huu Nguyen. Householder pseudo-rotation: A novel approach to activation editing in LLM s with direction-magnitude perspective. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Proceedings of the 2024 Conference on Empirical Methods in Nat...

    2024 doi

  27. [35]

    Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards

    Alexandre Rame, Guillaume Couairon, Corentin Dancette, Jean-Baptiste Gaya, Mustafa Shukor, Laure Soulier, and Matthieu Cord. Rewarded soups: towards pareto-optimal alignment by interpolating weights fine-tuned on diverse rewards. Advances in Neural Information Processing Syste...

    2024

  28. [36]

    Steering llama 2 via contrastive activation addition

    Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. Steering llama 2 via contrastive activation addition. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors, Proceedings of the 62nd Annual Meeting of the Association for Computationa...

    2024 doi

  29. [37]

    A consistent and efficient evaluation strategy for attribution methods

    Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consistent and efficient evaluation strategy for attribution methods. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, Proceedings of the ...

    2022

  30. [38]

    Are vision-language transformers learning multimodal representations? a probing perspective

    Emmanuelle Salin, Badreddine Farah, St \'e phane Ayache, and Benoit Favre. Are vision-language transformers learning multimodal representations? a probing perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11248--11257, 2022

    2022

  31. [39]

    Smith, and Simon Shaolei Du

    Ruizhe Shi, Yifang Chen, Yushi Hu, Alisa Liu, Hannaneh Hajishirzi, Noah A. Smith, and Simon Shaolei Du. Decoding-time language model alignment with multiple objectives. In ICML 2024 Workshop on Theoretical Foundations of Foundation Models, 2024. URL https://openreview.net/foru...

    2024

  32. [40]

    Smoothgrad: removing noise by adding noise, 2017

    Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. Smoothgrad: removing noise by adding noise, 2017. URL https://arxiv.org/abs/1706.03825

    2017 arXiv

  33. [41]

    Efficient open-set test time adaptation of vision language models

    Manogna Sreenivas and Soma Biswas. Efficient open-set test time adaptation of vision language models. In First Workshop on Scalable Optimization for Efficient and Adaptive Foundation Models, 2025. URL https://openreview.net/forum?id=ELR4ifkX3j

    2025

  34. [42]

    LVP runing: An effective yet simple language-guided vision token pruning approach for multi-modal large language models

    Yizheng Sun, Yanze Xin, Hao Li, Jingyuan Sun, Chenghua Lin, and Riza Batista-Navarro. LVP runing: An effective yet simple language-guided vision token pruning approach for multi-modal large language models. In Luis Chiruzzo, Alan Ritter, and Lu Wang, editors, Findings of the A...

    2025 doi

  35. [43]

    Aligning large multimodal models with factually augmented rlhf

    Zhiqing Sun, Sheng Shen, Shengcao Cao, Haotian Liu, Chunyuan Li, Yikang Shen, Chuang Gan, Liang-Yan Gui, Yu-Xiong Wang, Yiming Yang, et al. Aligning large multimodal models with factually augmented rlhf. arXiv preprint arXiv:2309.14525, 2023

    2023 arXiv

  36. [44]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Doina Precup and Yee Whye Teh, editors, Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 3319--3328. PML...

    2017

  37. [45]

    Gemini: A family of highly capable multimodal models, 2024

    Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. URL https://arxiv.org/abs/2312.11805

    2024 arXiv

  38. [46]

    Gemma 3 technical report, 2025

    Gemma Team et al. Gemma 3 technical report, 2025. URL https://arxiv.org/abs/2503.19786

    2025 arXiv

  39. [47]

    Qwen2.5: A party of foundation models, September 2024

    Qwen Team. Qwen2.5: A party of foundation models, September 2024. URL https://qwenlm.github.io/blog/qwen2.5/

    2024

  40. [48]

    Vazquez, Ulisse Mini, and Monte MacDiarmid

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid. Steering language models with activation engineering, 2024. URL https://arxiv.org/abs/2308.10248

    2024 arXiv

  41. [49]

    Contrastive region guidance: Improving grounding in vision-language models without training

    David Wan, Jaemin Cho, Elias Stengel-Eskin, and Mohit Bansal. Contrastive region guidance: Improving grounding in vision-language models without training. In European Conference on Computer Vision, pages 198--215. Springer, 2024

    2024

  42. [50]

    Ad-kd: Attribution-driven knowledge distillation for language model compression

    Siyue Wu, Hongzhan Chen, Xiaojun Quan, Qifan Wang, and Rui Wang. Ad-kd: Attribution-driven knowledge distillation for language model compression. arXiv preprint arXiv:2305.10010, 2023

    2023 arXiv

  43. [51]

    Nullu: Mitigating object hallucinations in large vision-language models via halluspace projection, 2025

    Le Yang, Ziwei Zheng, Boxu Chen, Zhengyu Zhao, Chenhao Lin, and Chao Shen. Nullu: Mitigating object hallucinations in large vision-language models via halluspace projection, 2025. URL https://arxiv.org/abs/2412.13817

    2025 arXiv

  44. [52]

    Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A m...

    2024

  45. [53]

    Spa-vl: A comprehensive safety preference alignment dataset for vision language model, 2025

    Yongting Zhang, Lu Chen, Guodong Zheng, Yifeng Gao, Rui Zheng, Jinlan Fu, Zhenfei Yin, Senjie Jin, Yu Qiao, Xuanjing Huang, Feng Zhao, Tao Gui, and Jing Shao. Spa-vl: A comprehensive safety preference alignment dataset for vision language model, 2025. URL https://arxiv.org/abs...

    2025 arXiv

  46. [54]

    Bayesian test-time adaptation for vision-language models, 2025

    Lihua Zhou, Mao Ye, Shuaifeng Li, Nianxin Li, Xiatian Zhu, Lei Deng, Hongbin Liu, and Zhen Lei. Bayesian test-time adaptation for vision-language models, 2025. URL https://arxiv.org/abs/2503.09248

    2025 arXiv

  47. [55]

    Representation engineering: A top-down approach to ai transparency

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, et al. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405, 2023

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.