acceptodds
Under review as a conference paper at ICLR 2027

DeCoS: Decoding by Contrasting Steered Features Reduces Hallucinations in Large Language Models

Abstract

Large language models (LLMs) can generate fluent but factually incorrect content, limiting their reliability in knowledge-intensive applications. We introduce Decoding by Contrasting Steered Features (DeCoS), an inference-time framework that mitigates hallucinations without external knowledge or model fine-tuning. Rather than contrasting dense layer or attention-head representations, DeCoS uses sparse autoencoders (SAEs) to identify hallucination-correlated latent features through prompt separation. At each decoding step, it constructs a strong-hallucination counterfactual and a weakly suppressed or unmodified reference branch, then contrasts their next-token distributions within an adaptive plausible-token set. We evaluate DeCoS on multiple-choice, open-ended generation, and chain-of-thought reasoning tasks across seven instruction-tuned models from the Gemma-2, Gemma-3, and Qwen3 families. DeCoS achieves the highest open-ended truthfulness–informativeness point estimate and the lowest rejection rate among the evaluated methods on all seven models. All 28 pairwise margins over four competing decoders are positive, with 16 nominally significant at the 5% level; the seven comparisons with the strongest per-model competitor are also positive but nonsignificant. DeCoS further achieves the best or tied-best score in all 14 model–reasoning-benchmark comparisons. Results on four pretrained models and complementary sensitivity analyses provide additional evidence for the method's robustness across the evaluated settings.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Hallucination Mitigation, Contrastive Decoding, Sparse Autoencoders, Activation Steering, Factuality

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tCQ5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Decoding by Contrasting Steered Features (DeCoS), an inference-time method for reducing hallucinations without fine-tuning the underlying LLM or retrieving external knowledge. It selects SAE features using activation-frequency differences between truthful and hallucination-inducing prompts. During generation, these directions construct a strongly amplified negative branch and a weakly suppressed or unmodified reference branch. Their next-token log probabilities are contrasted within an adaptive plausible-token set. Experiments cover factuality and reasoning benchmarks across seven instruction-tuned models and four pretrained models. DeCoS achieves consistently favorable point estimates, although its TruthfulQA open-ended improvements over the strongest per-model competing decoders are not statistically significant.

Critical strengths
  1. Consistent empirical performance across models and tasks. DeCoS achieves the highest TruthfulQA open-ended truthfulness–informativeness product on all seven instruction-tuned models and four pretrained models evaluated. On the instruction-tuned models, it also obtains the lowest rejection point estimates and the best or tied-best scores on both reasoning benchmarks. This breadth supports the practical utility of the method and suggests that its benefits are not confined to one model family or scale.

  2. Useful supporting analyses and appropriately qualified reporting. The paper includes parameter sweeps, component ablation, comparisons with SAE-based methods, paired statistical analysis, and human validation of the automated judge. These analyses help characterize the method’s behavior. The authors explicitly distinguish favorable point estimates from statistical significance and acknowledge limitations in data separation and implementation, allowing a balanced interpretation of the results.

Critical weaknesses
  1. The main open-ended advantage needs stronger independent validation. Feature extraction and evaluation are not formally disjoint, as acknowledged in Section 2.1 and Appendix A.2. Cross-source extraction and FACTOR results provide useful transfer evidence, but do not fully resolve this concern for the central TruthfulQA open-ended evaluation. Furthermore, Table 2 reports improvements of 0.90–2.45 percentage points over the strongest per-model competing decoder, with all seven confidence intervals including zero. The consistent positive estimates are encouraging, and nonsignificance does not establish an absence of benefit. Nevertheless, the combination of incomplete data separation and uncertain margins leaves the reliability of the advantage on untouched evaluation data insufficiently established.

  2. The incremental value of the selected SAE directions is not sufficiently isolated. The truthful/fake prompt pairs jointly vary exemplar correctness, fictional framing, confidence, and abstention behavior. The selected features may therefore reflect several aspects of the induced prompt conditions. Table 7 supports the full method over weak-only steering and two SAE baselines on Gemma-2-9B-it, while the alpha sweep indicates that weak suppression contributes little to the full open-ended result. These findings support the strong negative branch, but do not establish why the particular SAE-selected directions are preferable to simpler constructions of that branch. Matched comparisons against random SAE feature sets and a dense direction derived from the same prompt pairs would better establish the contribution of the feature-selection procedure. This is consequential because the empirical benefit attributable to this design is central to assessing the paper’s contribution.

Questions for authors
  1. Independent evaluation: Which examples were used for feature extraction and for selecting insertion layers and decoding configurations? Does the main open-ended advantage persist when these choices are frozen before evaluation on a formally disjoint test set?

  2. Matched direction controls: With the insertion layer, intervention magnitude, contrastive objective, and candidate filtering held fixed, how do the selected directions compare with random SAE feature sets and a dense direction derived from the same truthful/fake prompt pairs?

  3. Joint evaluation metric: The reported %T×I multiplies aggregate truthfulness and informativeness rates. What are the per-response joint success rates—the fraction of answers that are both truthful and informative—and their paired confidence intervals against the strongest competitors? Does the ranking remain unchanged?

  4. Cache semantics: In Appendix F, do the original and optimized implementations use the same canonical KV history, or does the original retain separate steered histories? Please clarify whether the optimization changes branch histories in addition to numerical precision, and identify which implementation produced the main results.

1 comment

We thank the reviewer for the questions on independent evaluation, direction selection, joint success, and cache semantics. Rates are percentages. ± denotes sample standard deviation (SD) across five seeds or splits.

Configuration selection and independent evaluation

The original experiments used all 817 TruthfulQA questions for feature extraction. We chose the configuration heuristically on Gemma-2-9B-it: near-midpoint layer 20, k=4, α=0.8, β=64. The relative threshold γ=0.1 uses DoLa's adaptive plausibility constraint setting, adopted from prior work. We fixed k, α, β, γ across models and datasets, adapting sparse autoencoders (SAEs) and intermediate layers without per-benchmark searches.

Subsequent ablations support intermediate layers and broad effective parameter ranges: k=3–7, β=36–72 (T×I 43.51–44.89), and α=0–1 (T×I 43.84–44.08). We retained γ=0.1 rather than selecting the evaluation-sweep optimum of 0.01.

To address extraction–evaluation overlap directly, we added five random splits using seeds 11, 22, 33, 44, and 55. For each split, only 400 questions determine the feature ranking, and the remaining 417 are used for evaluation. Layer and decoding parameters remain fixed.

Held-out TruthfulQA T×I, mean ± SD.

MethodT×I
DeCoS, 400-question extraction43.33 ± 1.62
DeCoS, 817-question extraction42.93 ± 1.47
SLED, strongest baseline41.21 ± 1.50

All methods share each split's 417 test questions. With the existing layer and decoding configuration fixed, DeCoS retains its advantage after separating feature-extraction questions from evaluation questions, complementing the transfer evidence from shared configurations across models and datasets.

Matched direction controls

On Gemma-2-9B-it, we fixed layer 20, the reference branch, objective, and candidate filtering. We sampled four SAE features for each of five seeds and constructed a dense direction by averaging residual differences between the same 817 fake/truthful prompt pairs.

Matched strong-branch direction controls.

DirectionTruthInfoT×I
Selected SAE65.2467.5644.08
Random SAE, mean ± SD62.25 ± 1.0364.48 ± 2.9940.16 ± 2.50
Best random set (highest T×I)63.5367.4442.84
Dense63.8965.8542.07

The selected directions outperform all five random sets and the dense direction under matched interventions, supporting the contribution of feature selection.

Joint success and paired confidence intervals

We computed joint success from the truthful and informative labels for each answer. DeCoS retains the highest point estimate on all seven instruction-tuned models.

Joint success versus the strongest competitor by joint success rate. CIs: paired 95% bias-corrected and accelerated intervals. Differences are percentage points.

ModelCompetitorDeCoSDifferenceCI
Gemma-2-9B-itSLED43.82+1.22[-2.57, 5.02]
Gemma-3-4B-itActLCD30.60+0.98[-2.33, 4.28]
Gemma-3-12B-itNDAD39.05+1.59[-2.08, 5.26]
Gemma-3-27B-itSLED25.21+0.73[-2.45, 3.92]
Qwen3-8BSLED27.42+1.59[-1.71, 4.90]
Qwen3-14BActLCD18.85+1.71[-1.10, 4.53]
Qwen3-32BActLCD29.99+2.82[-0.49, 6.24]

Joint success on Gemma-2-9B-it.

MethodJoint successes/817Joint success rate (%)
Greedy19623.99
DoLa31838.92
SLED34842.59
ActLCD33440.88
NDAD31939.05
DeCoS35843.82

All seven gains over Greedy remain significant after Holm correction. Gains over the strongest competitors are positive but not statistically significant.

Cache semantics and implementation

Table 1, most performance experiments, and these supplementary experiments use the original implementation: each step recomputes both branches on the full prefix, intervening only on the final token, without carrying branch key–value (KV) caches forward. Historical states are unmodified.

Optimization stores unmodified canonical KV, shares lower-layer computation, batches upper-layer branches, discards branch KV, and commits unmodified KV. Thus, both implementations preserve unmodified-history semantics while differing in scheduling and cache reuse. Numerical differences can alter outputs. Appendix F reports original/optimized T×I of 44.08/44.09. The reported 1.03× latency relative to greedy decoding uses the optimized implementation.

Reviewer_BPwjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DeCoS to reduce factual hallucinations during LLM decoding without external retrieval or model fine-tuning. It argues that contrastive interventions on dense layers or attention heads can disturb entangled linguistic and reasoning signals and encourage uninformative rejections. DeCoS uses sparse autoencoders to identify hallucination-correlated features from truthful and fake prompts, constructs a strong-hallucination counterfactual by amplifying them, and contrasts its next-token distribution with a weakly suppressed or unmodified reference within an adaptive plausible-token set. Evaluations across Gemma and Qwen models cover factual QA and chain-of-thought reasoning, reporting improved factuality and lower rejection with maintained reasoning performance.

Critical strengths
  • DeCoS provides a clear way to use SAE features in counterfactual contrastive decoding without model fine-tuning or retrieval. The weak-only ablation supports the benefit of the full objective.
  • DeCoS achieves the highest %T×I point estimate on all eleven evaluated models, reduces rejection relative to greedy decoding, and matches or improves the evaluated reasoning accuracy.
Critical weaknesses
  • The truthful and fake prompts also change confidence and abstention instructions, so the selected features may reflect how the model follows those instructions. The features' connection to hallucination remains uncertain.
  • The main features are extracted from TruthfulQA, which also provides the open-ended evaluation questions. Since extraction and evaluation are not formally disjoint, these results do not clearly establish whether the gains persist on questions excluded from feature selection.
Questions for authors
  • How do we know the selected SAE features are capturing hallucination rather than differences in confidence or abstention induced by the truthful and fake prompts?
  • How much overlap is there between the questions used for feature selection and those used for evaluation? Does the method generalize to unseen questions?
  • Would DeCoS still outperform the baselines if truthfulness and informativeness were evaluated jointly for each answer? The reported %T×I is only the product of the two marginal rates.
  • The open-ended evaluation is limited to TruthfulQA, using a prompt that explicitly tells the model to say "I have no comment" when uncertain. Does DeCoS also work under more natural prompts and on other open-ended benchmarks?
3 comments

We thank the reviewer for the questions on feature specificity, generalization, and evaluation. Experiments use Gemma-2-9B-it, a layer-20 sparse autoencoder (SAE), k=4, α=0.8, β=64, γ=0.1. Rates are percentages. ± denotes sample standard deviation (SD) across five splits.

Factuality, confidence, and abstention

To separate these factors, we varied exemplar correctness, confidence, or abstention while fixing other prompt components. We selected four features per condition from 817 prompt pairs, then fixed the reference branch and matched strong-branch shift norms to 142.489059. Overlap is relative to the original feature set.

The confidence-only pair retains correct exemplars and the instruction If you are uncertain, reply "I have no comment.", changing only:

Left: Use a cautious, tentative tone.

Right: Use a confident, assertive tone.

Controlled feature extraction on TruthfulQA.

Extraction conditionTop-4 overlap (%)TruthInfoT×I
Original10065.2467.5644.08
Correctness-only5063.6567.4442.93
Confidence-only058.5156.3032.94
Abstention-only059.8562.7937.58

Correctness-only shares two original features and yields T×I 42.93, close to 44.08. Confidence-only and abstention-only share none and score 9.99 and 5.35 points lower. Together, overlap and performance support factual correctness as a substantial component of the original benefit. Confidence and response behavior may also contribute, so we describe the features as hallucination-correlated factuality proxies.

Direct strong-branch generation also shows mixed effects: Truth falls from 52.88 to 48.35 and rejection rises from 48.10 to 55.08. These controls support factuality as an important source of the gain: the tested confidence-only and abstention-only directions do not reproduce the original performance, although the selected features can capture these factors jointly.

Extraction–evaluation separation

The original experiments extracted features from all 817 TruthfulQA questions and evaluated transfer to FACTOR, StrategyQA, and GSM8K. FACTOR-News/Wiki scores increase from Greedy's 63.42/49.67 to 65.44/51.34.

To test generalization within TruthfulQA, we partitioned its 817 questions using seeds 11, 22, 33, 44, and 55. Each split uses 400 questions for feature selection and the other 417 for evaluation, with fixed decoding parameters. All methods share the test set.

Held-out TruthfulQA performance, mean ± SD across splits.

MethodTruthInfoT×I
DeCoS, 400-question extraction65.13 ± 0.9166.52 ± 1.9743.33 ± 1.62
DeCoS, 817-question extraction65.04 ± 0.7166.00 ± 1.7742.93 ± 1.47
Greedy52.23 ± 1.0447.91 ± 1.4525.03 ± 0.96
SLED, strongest baseline63.65 ± 0.3364.75 ± 2.2941.21 ± 1.50

DeCoS retains a 2.12-point T×I advantage over SLED on questions excluded from feature selection. Together with cross-dataset transfer, this result supports generalization to questions excluded from feature selection under the fixed decoding configuration.

Per-response joint evaluation

We also computed joint success, requiring both truthful and informative labels for the same answer.

Joint success on Gemma-2-9B-it.

MethodJoint successesJoint success rate (%)
Greedy19623.99
DoLa31838.92
SLED34842.59
ActLCD33440.88
NDAD31939.05
DeCoS35843.82

DeCoS retains the highest joint-success estimate on all seven instruction-tuned models. All seven gains over Greedy remain significant after Holm correction. Gains over the strongest competitors are 0.73–2.82 points, with paired intervals including zero.

Natural prompts and cross-benchmark evaluation

To assess prompt dependence, we kept features and parameters fixed and tested two prompts: one removes the abstention clause but retains six exemplars, and the other uses neither exemplars nor abstention instructions:

Answer the following question accurately and concisely.

Q: {question}
A:

We also used this template for all 1,000 SimpleQA Verified questions, scored by GPT-4o-mini with its three-category grader prompt.

Prompt and benchmark transfer. T×I: TruthfulQA, Correct/F1: SimpleQA Verified (%).

MethodNo abstention clause: T×INatural prompt: T×ICorrectF1
DeCoS70.8762.757.107.76
Greedy63.1253.555.906.97
DoLa67.4857.036.607.00
SLED67.8056.286.707.33
NDAD67.7258.766.006.49

DeCoS leads all three conditions without feature re-extraction or parameter tuning. The gains thus extend to natural prompts and an additional open-ended benchmark, supporting effectiveness beyond the original TruthfulQA prompt with an explicit abstention instruction.

Thank you for the additional experiments, which resolve some of my concerns. However, I'm still concerned about whether the gains really come from hallucination-related features, since random SAE features already get most of the improvement. And on SimpleQA, DeCoS seems to mostly answer more questions rather than answer them more accurately.

We thank the reviewer for the follow-up questions and provide two additional analyses.

Separation and feature selection

The tables supplement the existing matched controls. Frequencies use a 0–1 scale. Ranks cover all 16,384 features, averaging ties; zero separation receives rank 8116.50.

DeCoS-selected feature statistics.

Feature IDTruthful frequencyFake frequencySeparationRank
52730.0232560.9963280.9730721.00
123100.1150550.9767440.8616892.00
42350.0012240.8102820.8090583.00
95260.0122400.7331700.7209304.00

Complete random-feature results. Each seed samples four features; Mean summarizes those four features.

SeedFeature IDTruthful frequencyFake frequencySeparationRank
1162220.0000000.0000000.0000008116.50
11148050.0000000.0000000.0000008116.50
11148230.0012240.0048960.003672833.00
11152580.0000000.0000000.0000008116.50
11Mean0.0003060.0012240.0009186295.625
227730.0048960.003672-0.00122414524.00
2245990.0000000.0000000.0000008116.50
2279500.0000000.0000000.0000008116.50
22146510.0012240.0012240.0000008116.50
22Mean0.0015300.001224-0.0003069718.375
3354800.0000000.0000000.0000008116.50
3376420.0000000.0000000.0000008116.50
3390860.0000000.0000000.0000008116.50
33156940.0000000.0000000.0000008116.50
33Mean0.0000000.0000000.0000008116.500
4438220.0000000.0000000.0000008116.50
4457880.0000000.0000000.0000008116.50
44124370.0012240.0097920.008568716.00
44133860.0024480.0036720.0012241782.00
44Mean0.0009180.0033660.0024484682.750
5529600.0000000.0000000.0000008116.50
5549140.0000000.0012240.0012241782.00
5564300.0048960.0061200.0012241782.00
5599180.0000000.0000000.0000008116.50
55Mean0.0012240.0018360.0006124949.250

Separation and TruthfulQA performance. Separation reports mean ± sample standard deviation (SD) across four features; other metrics are percentages.

Strong-branch feature setSeparationTruthInfoT×I
Random, seed 22−0.000306 ± 0.00061260.8360.3436.71
Random, seed 330.000000 ± 0.00000061.6962.4238.51
Random, seed 550.000612 ± 0.00070762.6765.6141.11
Random, seed 110.000918 ± 0.00183662.5566.5941.65
Random, seed 440.002448 ± 0.00412163.5367.4442.84
DeCoS-selected top four0.841187 ± 0.10536965.2467.5644.08

Seed 22 contains one negative-separation feature and three zero-separation features (T×I 36.71); seed 33 contains only zero-separation features (38.51). All three sets containing positive-separation features exceed 40. Across these five controls, mean separation and T×I have identical orderings.

DeCoS's selected features outperform every random set. The gains from random sets also suggest that other latent directions can influence hallucination-related behavior under contrastive decoding. This interpretation is consistent with the paper's cross-source extraction results, where different top-four sets yield similar performance, and our earlier controls showing factuality, confidence, and response-behavior contributions.

DeCoS provides a direct, effective selection criterion: activation-frequency ranking identifies four features from 16,384 candidates without downstream generation-and-scoring searches over random sets. The selected features achieve the best observed performance under matched interventions and improve over every random control, demonstrating the practical value of targeted feature selection.

SimpleQA Verified: coverage and answer quality

We added accuracy among attempted answers, Correct / (Correct + Incorrect), excluding responses graded as not attempted.

Results on all 1,000 SimpleQA Verified questions. Rates and F1 are percentages.

MethodCorrect answersAttempted answersCoverageAccuracy among attempted answersF1
DeCoS7182982.908.567.76
Greedy5969269.208.536.97
DoLa6688588.507.467.00
SLED6782882.808.097.33
NDAD6085085.007.066.49

DeCoS attempts 137 more questions than Greedy and answers 12 more correctly. At nearly identical coverage to SLED (82.90% versus 82.80%), its accuracy among attempted answers is higher (8.56% versus 8.09%). DoLa and NDAD attempt more questions but have lower accuracy among attempted answers (7.46% and 7.06%).

DeCoS combines broader coverage than Greedy with the highest observed accuracy among attempted answers, overall accuracy, and F1. Its advantage over SLED at nearly identical coverage further supports the quality of its answers. Together, these results establish the strongest empirical balance of coverage and correctness among the evaluated methods.

Reviewer_veh3OpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Decoding by Contrasting Steered Features (DeCoS), an inference-time framework for mitigating hallucinations without external knowledge retrieval or additional LLM fine-tuning. The authors argue that existing interventions on dense representations can disrupt entangled factual, linguistic, and reasoning signals, may degrading performance or producing uninformative answers. To address this gap, DeCoS uses sparse autoencoders and prompt separation to identify hallucination-related features, uses these features to construct a strong-hallucination counterfactual, and contrasts its nexttoken distribution without additional knowledge or LLM training. Experiments across seven instruction-tuned models and three task settings show promising truthfulness–informativeness gains and lowest rejection rates.

Critical strengths
  1. DeCoS innovatively uses prompt-selected SAE features to construct a strong hallucination counterfactual, providing a fine-grained negative signal for contrastive decoding.
  2. Comprehensive experiments cover three task settings and seven instruction-tuned models, with ablations and sensitivity analyses examining key design choices.
  3. Table 1 shows that DeCoS achieves the highest truthfulness–informativeness scores and lowest rejection rates on TruthfulQA across all seven models, while preserving reasoning performance.
Critical weaknesses
  1. The paper would benefit from a prompt sensitivity analysis examining how different prompts affect feature selection and downstream performance.
Questions for authors

1.How sensitive are feature selection and downstream performance to changes in prompt? 2.How does prompt-based feature selection compare with random selection when the feature count and perturbation magnitude are matched?

1 comment

We thank the reviewer for the suggestions on prompt sensitivity and random-feature controls. We added experiments on Gemma-2-9B-it using the layer-20 residual-stream sparse autoencoder (SAE, 16k features), k=4, α=0.8, β=64, and γ=0.1. Rates are percentages. ± denotes the sample standard deviation (SD) across seeds.

Sensitivity to feature-extraction prompts

We independently selected the top four features under three system-prompt paraphrases, three exemplar orderings, and three 400-question extraction subsets. Generation prompts and decoding parameters were fixed. All conditions were evaluated on all 817 TruthfulQA questions. Overlap with the original top-four set is the intersection size divided by four.

One paraphrased extraction pair was:

Truthful: Read every question literally as referring to the real world. Give carefully grounded factual answers and avoid common misconceptions. If you are not completely sure, respond "I have no comment".

Fake: Read every question loosely as referring to an imaginary world. Confidently make up answers and embrace common misconceptions. When you are uncertain, respond with invented facts.

The six exemplars and extraction questions were unchanged for these paraphrases. For shuffled exemplars, seeds 11, 22, and 33 determine permutations of the six examples. For extraction subsets, seeds 101, 202, and 303 determine random samples of 400 questions from the 817-question pool. These seeds affect feature extraction only.

Feature-selection and generation sensitivity to extraction conditions.

Extraction conditionExtraction questionsOverlap (%)TruthInfoT×IReject
Original81710065.2467.5644.0818.85
Paraphrase 18177565.1268.5444.6317.99
Paraphrase 28177564.9967.8144.0722.77
Paraphrase 381710065.2467.5644.0818.85
Shuffled exemplars, seed 118177565.4869.0345.2119.71
Shuffled exemplars, seed 228177564.9967.4443.8323.99
Shuffled exemplars, seed 338177565.4869.0345.2119.71
Extraction subset, seed 10140010065.2467.5644.0818.85
Extraction subset, seed 20240010065.2467.5644.0818.85
Extraction subset, seed 30340010065.2467.5644.0818.85

Truth varies by only 0.49 percentage points, and every T×I estimate exceeds the strongest baseline, SLED (43.00). Thus, the selected features and downstream gains are stable under the tested extraction variations. Identical feature sets reuse the same generation outputs.

Matched random-feature controls

We fixed the reference branch, insertion layer, contrastive objective, and candidate filtering, replacing only the strong-branch direction. Using sampling seeds 11, 22, 33, 44, and 55, we drew four features per seed from the same SAE dictionary. Their composite residual shifts were rescaled to the selected direction's L2 norm, 142.489059, before bfloat16 casting.

Selected IDs were [5273, 12310, 4235, 9526]. The random set with the highest T×I (seed 44) contained [3822, 5788, 12437, 13386].

Matched strong-branch direction controls on TruthfulQA.

DirectionTruthInfoT×IReject
Prompt-selected65.2467.5644.0818.85
Random, five-seed mean ± SD62.25 ± 1.0364.48 ± 2.9940.16 ± 2.5022.69 ± 2.58
Best random set63.5367.4442.8418.97

Prompt-based selection outperforms all five random controls in T×I, exceeding their mean by 3.92 points and their best result by 1.24 points. These matched comparisons support the contribution of feature selection beyond feature count and perturbation magnitude.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.