PaperlayerView exact revision on arXiv

Source-linked AI summary

Self-Improving LLM Agents at Test-Time

Emre Can Acikgoz, Cheng Qian, Heng Ji, Dilek Hakkani-Tür, Gokhan Tur

arXiv:2510.07841v1cs.LGcs.AIcs.CL

TL;DR

Large-scale LM fine-tuning is costly, and existing approaches may treat samples as equally informative. TT-SI selects uncertain test instances, generates similar examples, and adapts at test time; across four benchmarks, it reports a 5.48% average absolute accuracy gain for direct inference.

  • Problem

    Large-scale LM fine-tuning is costly, while existing approaches treat samples as equally informative despite distribution shift and redundancy.

    Source: §2.1 · p. 3

  • Method

    TT-SI identifies uncertain test instances, synthesizes distributionally similar examples, and uses them for temporary test-time updates.

    Source: §1 · p. 2, §3.2.1 · p. 6

  • Results

    5.48% average absolute accuracy gain for direct inference across four benchmarks is reported for TT-SI using one generated training instance.

    Source: §4.1 · p. 8

  • Takeaways & Limitations

    The findings support the feasibility of agent self-improvement during inference with only one training instance.

    Source: §4.1 · p. 8, §5 · p. 10

    TT-SI cannot recover task knowledge absent from the pretrained model, and its best performance is sensitive to the uncertainty threshold.

    Source: §5 · p. 10

Abstract

from arXiv · show

One paradigm of language model (LM) fine-tuning relies on creating large training datasets, under the assumption that high quantity and diversity will enable models to generalize to novel tasks after post-training. In practice, gathering large sets of data is inefficient, and training on them is prohibitively expensive; worse, there is no guarantee that the resulting model will handle complex scenarios or generalize better. Moreover, existing techniques rarely assess whether a training sample provides novel information or is redundant with the knowledge already acquired by the model, resulting in unnecessary costs. In this work, we explore a new test-time self-improvement method to create more effective and generalizable agentic LMs on-the-fly. The proposed algorithm can be summarized in three steps: (i) first it identifies the samples that model struggles with (self-awareness), (ii) then generates similar examples from detected uncertain samples (self-data augmentation), and (iii) uses these newly generated samples at test-time fine-tuning (self-improvement). We study two variants of this approach: Test-Time Self-Improvement (TT-SI), where the same model generates additional training examples from its own uncertain cases and then learns from them, and contrast this approach with Test-Time Distillation (TT-D), where a stronger model generates similar examples for uncertain cases, enabling student to adapt using distilled supervision. Empirical evaluations across different agent benchmarks demonstrate that TT-SI improves the performance with +5.48% absolute accuracy gain on average across all benchmarks and surpasses other standard learning methods, yet using 68x less training samples. Our findings highlight the promise of TT-SI, demonstrating the potential of self-improvement algorithms at test-time as a new paradigm for building more capable agents toward self-evolution.

1 INTRODUCTION

The introduction argues that costly, exhaustive fine-tuning can waste effort on redundant samples without ensuring generalization, and proposes targeted, human-inspired adaptation at test time. It reports benchmark gains from uncertainty-guided self-improvement using minimal synthesized data.

1 INTRODUCTION

  • Standard agent fine-tuning requires costly large datasets, processes potentially redundant examples, and offers no guarantee of effective performance or generalization.The introduction motivates targeted learning by contrasting broad practice with identifying knowledge gaps, generating relevant questions, and practicing them.Source: §1 · p. 2
  • The proposed three-stage method identifies uncertain samples, generates similar training instances for retained test cases, and performs temporary test-time updates.An uncertainty estimator discards mastered or redundant samples; data synthesis and test-time fine-tuning then support targeted adaptation.Source: §1 · pp. 2–3
  • TT-SI learns from its own synthesized samples using PEFT, while TT-D uses samples synthesized by a more capable teacher for supervision.The introduction says TT-D extends gains in complex, context-heavy scenarios such as multi-turn conversations.Source: §1 · p. 2
  • +5.84% on ToolAlpaca, +6.05% on NexusRaven, +5.76% on SealTool, and +4.26% on API-Bank are TT-SI’s absolute accuracy gains with one synthesized training instance per test case.The introduction also reports that TT-SI surpasses SFT accuracy on SealTool with 68× fewer samples.Source: §1 · p. 2
  • When training is infeasible, test-time ICL offers a fast, training-free alternative that outperforms other standard learning methods in similar conditions.The paper also describes a systematic study of both variants and factors including uncertain-sample impact, test-time learning method, and generated-sample scaling.Source: §1 · pp. 2–3

2 PRELIMINARIES

The preliminaries explain why inductive fine-tuning can be costly and unreliable under distribution shift, then motivate adapting at test time using informative examples. They situate TT-SI in prior test-time training and self-improvement work, framing its mechanism as eliciting latent model knowledge through sample-specific adaptation.

2.1 FUNDAMENTAL ISSUES IN INDUCTIVE FINE-TUNING

  • Inductive fine-tuning assumes large, diverse training sets will generalize, but distribution shift, data costs, redundancy, forgetting, and model churn undermine that approach.The paper notes that the number of truly effective samples can be much smaller than the training set, while repeated retraining is needed to use newer base models.Source: §2.1 · p. 3
  • These limitations motivate transductive, local adaptation that identifies and trains on informative samples from the test distribution.Source: §2.1 · p. 4

2.2 TEST-TIME TRAINING

  • Prior test-time training includes self-supervised adaptation, nearest-neighbor fine-tuning, and exemplar transformations, while this work generates and filters signals for agent benchmarks.The paper characterizes its approach as the first language-generation-based test-time fine-tuning method for LLM-based agents.Source: §2.2 · p. 4

2.3 SELF-IMPROVEMENT IN LLMS AND HOW IT WORKS

  • Self-improvement is framed as sharpening outputs using latent knowledge already in model weights, rather than creating knowledge without new information.The account attributes sharpening to favoring high-confidence predictions aligned with internal self-evaluation criteria (Huang et al., 2025).Source: §2.3 · p. 4
  • TT-SI detects uncertain inputs, synthesizes similar training instances, and performs targeted temporary fine-tuning where adaptation is expected to yield the greatest marginal benefit.The method uses an uncertainty estimator, data synthesis function, and test-time fine-tuning to learn sample-specific temporary parameters.Source: §2.3 · p. 4

3 METHOD

The method adapts agents at test time by detecting uncertain inputs, generating targeted synthetic training data, and briefly fine-tuning the model for the current query. Adaptation focuses on challenging cases while restoring the original model parameters afterward.

3 METHOD

  • The framework combines uncertainty detection, targeted data synthesis, and test-time fine-tuning to adapt agents on challenging instances as they arise.The three components are Uncertainty Estimator (H), Data Synthesis Function (G), and Test-Time Fine-tuning (T).Source: §3 · p. 5

3.1 SELF-AWARE SAMPLE SELECTION AT TEST TIME

  • Inputs are flagged as uncertain without ground-truth labels when their confidence or softmax-difference falls below a user-defined threshold.Confidence is estimated over available actions using negative log-likelihood and Relative Softmax Scoring; uncertainty is the difference between the top two RSS scores.Source: §3.1 · pp. 5–6

3.2 DATA GENERATION STRATEGIES

  • For each uncertain input, the method triggers on-the-fly synthesis of a focused, temporary dataset to support localized adaptation.The synthesis process is specific to the uncertain query and is intended to address the challenge the model encounters.Source: §3.2 · p. 6

3.2.1 DATA SYNTHESIS METHOD

  • The model uses the unlabeled uncertain input as a seed to generate K semantically similar input-output pairs with slight variations.K is user-defined, and the generated examples are intended to be plausible samples from the underlying distribution around the model’s uncertainty region.Source: §3.2.1 · pp. 6–7

3.3 TEST-TIME FINE-TUNING

  • For each uncertain query, the model is temporarily fine-tuned on its synthetic dataset, then restored to its original parameters before the next input.The update minimizes a task-specific loss; the method uses LoRA for computational efficiency during inference-time updates.Source: §3.3 · p. 7

4 RESULTS

Across four agent benchmarks, test-time self-improvement improves accuracy with few samples; further analyses examine data efficiency, inference-time alternatives, uncertainty thresholds, and model scale.

4 RESULTS

  • The experiments used Qwen2.5-1.5B-Instruct as the main model and repeated all experiments five times, reporting averages.Evaluation covered four complementary agent benchmarks; Qwen2.5-7B-Instruct was included in scaling analyses.Source: §4 · p. 7

4.1 MAIN RESULTS

  • TT-SI gained 5.48% in direct inference, 3.85% in majority voting, and 3.46% in Pass@5 on average; TT-D added 0.94%, 1.68%, and 2.65%, respectively.The main results evaluate NexusRaven, SealTool, API-Bank, and ToolAlpaca; TT-SI uses one generated sample per uncertain case.Source: Table 1 · p. 8, §4.1 · p. 8
  • TT-SI with SFT reached 72.43% accuracy versus 70.20% for inductive SFT using roughly 68× fewer samples: 190 uncertain cases rather than about 13,000.The comparison is on SealTool.Source: §4.1 · p. 8, Figure 2 · p. 9
  • On SealTool, TT-SI with ICL improved accuracy from 66.37% to 68.36%, exceeding the standard ICL baseline of 67.74% without fine-tuning.Generated examples are inserted into the prompt as demonstrations.Source: §4.1 · p. 8, Figure 2 · p. 9

4.2 ABLATION STUDIES AND ANALYSIS

  • On OOD xLAM data, TT-SI outperformed SFT at every tested scale, with larger gains as uncertain examples increased; TT-SI with ICL also surpassed SFT at matching data amounts.The tested scales were 1, 2, 4, and 8 samples.Source: §4.2 · p. 9, Figure 2 · p. 9
  • At τ = 0.95, TT-SI achieved 72.43% accuracy with 190 updates, while τ = 0.35 yielded 68.10% accuracy, FPR 0.09, and TPR 0.42.Selecting all 294 samples at high τ produced 73.47% accuracy but substantial overhead; all tested τ values exceeded the 66.37% base accuracy.Source: §4.2 · p. 9, Table 2 · p. 9
  • TT-SI improved accuracy by 5.76% on Qwen2.5-1.5B-Instruct and 3.02% on Qwen2.5-7B-Instruct, with a larger relative boost for the smaller model.The reported accuracy pairs were 66.67→72.43 and 80.95→83.97, respectively.Source: §4.2 · p. 9, Figure 3 · p. 9
  • On SealTool, adapting uncertain samples achieved 72.43% accuracy versus 70.07% for certain samples; adapting all samples added only marginal gain while processing 104 more cases.The results support targeting uncertain inputs to focus adaptation where the model struggles and reduce computational cost.Source: §4.1 · p. 8, §4.2 · p. 10, Figure 4 · p. 10

5 DISCUSSIONS

The discussion presents TT-SI as a modular test-time learning approach that uses uncertainty to trigger targeted data synthesis and lightweight fine-tuning, while identifying limits and directions for extending it.

Conclusion

  • TT-SI uses uncertainty estimation, targeted training-instance synthesis, and test-time fine-tuning to improve agent performance during inference, even when training with one instance.Across benchmarks, it consistently improves test-time performance and achieves higher accuracy and efficiency than standard inductive learning baselines.Source: §5 · p. 10

Impact Statement

  • The contribution is a modular algorithm integrating self-awareness, self-augmentation, and self-improvement, rather than a particular uncertainty metric or data generator.Stronger update rules, improved uncertainty quantification, and better data generation can replace the current components.Source: §5 · p. 10

Limitations

  • TT-SI’s threshold τ affects best performance, despite consistent gains across tested values; autonomously learning the threshold remains an open challenge (Bakman et al., 2025).Source: §5 · p. 10
  • TT-SI is bounded by the base model’s knowledge: it cannot recover absent pretrained knowledge, for which retrieval or search may be necessary.Source: §5 · p. 10

Future Work

REPRODUCIBILITY STATEMENT … C PREVIOUS WORK ON LARGE LANGUAGE MODELS (LLMS) AND AGENT FINE-TUNING

The appendix distinguishes task-specific self-improvement from broader self-evolution, uses learning analogies to motivate uncertainty-driven training, and situates TT-SI against agent fine-tuning while documenting experimental reproducibility.

REPRODUCIBILITY STATEMENT

  • Experiments, including baselines, were repeated with five random seeds and run on one NVIDIA A40 GPU, using public agent benchmarks and reporting averages with standard deviations when appropriate.The algorithm is provided in Algorithm 1; hyperparameters and evaluation metrics are in Section G.Source: REPRODUCIBILITY STATEMENT · p. 11

A TERMINOLOGY: SELF-IMPROVING AND SELF-EVOLVING AGENTS

  • Self-improving agents refine task performance using internal evaluations without adding external knowledge (Huang et al., 2025), whereas self-evolving agents adapt more broadly by changing parameters, memory, tools, prompts, or architecture (Gao et al., 2025).The distinction is between bounded task-specific refinement and continuous adaptation for generalization across dynamic environments.Source: A TERMINOLOGY: SELF-IMPROVING AND… · p. 18

B OTHER EXAMPLES FROM SELF-REGULATED LEARNING

  • TT-SI draws on self-regulated learning: students identify uncertainty, seek related examples, and use them to address knowledge gaps.Unlike human learning, the method automatically generates semantically similar but varied examples without external supervision.Source: B OTHER EXAMPLES FROM SELF-REGULATED… · p. 18
  • Unlike classical active learning, which queries an oracle or annotator for informative examples, TT-SI generates targeted examples automatically as a cost-efficient alternative to data collection.The examples are semantically similar but slightly varied problem instances.Source: B OTHER EXAMPLES FROM SELF-REGULATED… · p. 18
  • A running-back training analogy illustrates the method’s sequence: diagnose the limiting factor, target relevant drills, and refine performance through repeated exposure rather than untargeted training.The passage contrasts speed- and direction-specific drills with nonspecific hypertrophy work that can increase fatigue and impair sprint performance.Source: B OTHER EXAMPLES FROM SELF-REGULATED… · p. 18

C PREVIOUS WORK ON LARGE LANGUAGE MODELS (LLMS) AND AGENT FINE-TUNING

  • Prior agent fine-tuning uses curated or synthesized task corpora (Ouyang et al., 2022; Wang et al., 2023), but can cause catastrophic forgetting, incur costly data generation, and fail to improve strong base models consistently.Agent work targets tool and API use (Yao et al., 2023; Patil et al., 2024) and generalizability (Zeng et al., 2024; Mitra et al., 2024; Chen et al., 2024b).Source: C PREVIOUS WORK ON LARGE LANGUAGE… · p. 19
  • The paper presents its work as the first application of test-time training to LLM agents, using temporary inference-time parameter updates and selecting informative samples to reduce reliance on large offline datasets and redundant training.The passage states that temporary updates avoid catastrophic forgetting.Source: C PREVIOUS WORK ON LARGE LANGUAGE… · p. 19

D UNCERTAINTY ESTIMATION RESULTS

The RSS-based uncertainty estimator separates correct from incorrect predictions more clearly than PPL on SealTool. Its threshold controls how many cases are flagged for adaptation, trading error coverage against unnecessary intervention.

D UNCERTAINTY ESTIMATION RESULTS

  • RSS shows clearer separation of correct and incorrect predictions than PPL on SealTool, with correct cases tending toward larger score differences and incorrect cases toward smaller ones.The comparison uses Qwen-2.5-1.5B-Instruct; PPL’s correct and incorrect score distributions overlap heavily, while RSS supports more reliable uncertainty filtering.Source: Figure 5 · p. 19, D UNCERTAINTY ESTIMATION RESULTS · p. 19
  • At τ = 0.95, Qwen-2.5-1.5B-Instruct reaches 96% TPR, 53% FPR, and 46.0% Youden’s J; τ = 0.95 is the adopted default.Across thresholds from 0.35 to 0.95, TPR rises from 42% to 96% and FPR from 9% to 53%.Source: D UNCERTAINTY ESTIMATION RESULTS · p. 20
  • At τ = 0.95, Qwen-2.5-7B-Instruct reaches 99% TPR and 49% FPR, with a Youden’s index of 50%.Across thresholds from 0.35 to 0.95, TPR rises from 39% to 99% and FPR from 15% to 49%.Source: D UNCERTAINTY ESTIMATION RESULTS · p. 20
  • Predictions are flagged as uncertain when the top-two RSS softmax difference u(x_i) is below τ; raising τ increases both TPR and FPR, affecting coverage and adaptation efficiency.The paper uses τ = 0.95 by default and notes that threshold choice affects computational cost.Source: D UNCERTAINTY ESTIMATION RESULTS · p. 20
  • 96.10% TPR, 53.46% FPR, and 42.64% Youden’s J make H strongest overall on SealTool, with J exceeding PPL’s 18.89% and H also achieving the best F1 score.The evaluation compares H against Random, Trivial, and PPL on Qwen-2.5-1.5B-Instruct.Source: D UNCERTAINTY ESTIMATION RESULTS · pp. 20–21

E DATA GENERATION DETAILS

For each uncertain input, the agent generates label-free, seed-based variants on the fly to support targeted adaptation to queries it struggles with.

E DATA GENERATION DETAILS

  • For each uncertain input x_i, the agent itself performs data synthesis using a hand-crafted prompt, the input without its label, and a specified sample count K.The prompt is shown in Figure 8.Source: E DATA GENERATION DETAILS · p. 21
  • The model generates K new input-output pairs whose variants preserve the seed’s core meaning and task relevance while adding controlled surface-level variation.The seed-based process is inspired by self-instruction methodologies (Wang et al., 2023).Source: E DATA GENERATION DETAILS · p. 21
  • Generating data on the fly for each uncertain instance enables targeted, timely adaptation to the kinds of queries the model struggles with as they arise.Source: E DATA GENERATION DETAILS · p. 21

E.1 EXPERIMENTAL RESULTS

For each uncertain query, Qwen-2.5-1.5B-Instruct creates one answer-bearing synthetic example without seeing the gold label for immediate inference-time adaptation. In SealTool, generated queries cluster near the uncertain input in UMAP space.

E.1 EXPERIMENTAL RESULTS

  • For each uncertain input, Qwen-2.5-1.5B-Instruct generates exactly one revised input and answer without the gold label, forming a temporary dataset for immediate adaptation.The generator receives only the instruction and query of the uncertain input.Source: §E.1 · p. 21
  • Generated queries form a compact cluster close to the uncertain SealTool input, indicating semantic alignment in the visualization.Figure 7 shows ten randomly sampled synthetic queries near the uncertain input.Source: §E.1 · p. 21, Figure 7 · p. 21
  • The SealTool visualization embeds instruction-query-output text with 768-dimensional Sentence-BERT vectors, then projects them using UMAP with 15 neighbors and minimum distance 0.1.The embedding combines system prompt, user query, and output, or equivalent instruction-input-output triples for generated data.Source: §E.1 · pp. 21–22

F CHEATING EXPERIMENTS AND TT-SI COMPARISON

On SealTool, TT-SI approaches the performance of cheating TTT despite using similar generated samples rather than test-set ground truth. The experiment also shows that update-based methods do not reach perfect accuracy after one epoch, unlike ICL with test examples in its prompt.

F CHEATING EXPERIMENTS AND TT-SI COMPARISON

  • The cheating experiment trains baselines on test data as an unrealistic upper bound, while TT-SI and TT-D use their standard configurations without test-set training.Using Qwen-2.5-1.5B-Instruct, the comparison is on SealTool; Figure 9's left four bars are cheating baselines and right bars are TT-SI and TT-D.Source: F CHEATING EXPERIMENTS AND TT-SI… · p. 22
  • 78.89% for cheating TTT versus 72.43% for TT-SI suggests similar generated samples can shift the model toward uncertain test samples without exact ground-truth answers.The authors interpret this as increased confidence on samples overlooked by the base model after temporary updates.Source: F CHEATING EXPERIMENTS AND TT-SI… · p. 22
  • 100% accuracy for ICL with test samples in the prompt contrasts with update-based SFT and TTT, neither of which reaches 100% after one training epoch.SFT reaches 96.60% after 10 epochs on the 294-sample test set.Source: F CHEATING EXPERIMENTS AND TT-SI… · p. 22

G IMPLEMENTATION DETAILS OF TT-SI … G.3 TRAINING

The implementation uses Qwen2.5-1.5B-Instruct for uncertainty estimation, sample generation, and test-time fine-tuning, with TT-D substituting GPT-5-mini for generation. The sections detail the uncertainty, generation, and LoRA training procedures and their costs.

G IMPLEMENTATION DETAILS OF TT-SI

  • The implementation uses Qwen2.5-1.5B-Instruct for uncertainty estimation, data synthesis, and test-time fine-tuning on a single NVIDIA A40 GPU.The components correspond to H, G, and T, described in Sections 3.1–3.3.Source: G IMPLEMENTATION DETAILS OF TT-SI · p. 23

G.1 UNCERTAINTY ESTIMATION

  • 0.87 seconds is the average time to label a SealTool sample uncertain using candidate-function confidence scores normalized with RSS.The procedure queries with the test instruction, extracts function names using regex, and does not apply temperature scaling.Source: §G.1 · p. 23

G.2 DATA GENERATION

  • TT-SI generates K labeled variations with Qwen2.5-1.5B-Instruct, while TT-D uses GPT-5-mini to generate samples from uncertain examples.The seed sample is included in the generation prompt, which requests slight variations rather than an identical sample.Source: §G.2 · p. 23
  • 3.45 seconds is the average TT-SI generation time per sample; JSON parsing errors allow up to five retries and usually resolve by the second or third attempt.TT-SI generation uses vLLM with temperature 0.7 and maximum length 32768; GPT-5-mini uses the standard OpenAI API, with negligible cost per experiment.Source: §G.2 · p. 23

G.3 TRAINING

  • Training uses LoRA with rank 8 and α 16 on all linear layers, for five epochs at learning rate 1.0 × 10^-4, warm-up ratio 0.03, and batch size 1.Fine-tuning uses LLaMA-Factory and a cosine scheduler; other LLaMA-Factory and HuggingFace settings remain at default.Source: §G.3 · p. 24
  • 2.05 seconds is the average time to train on one sample; training uses Alpaca-style data with loss computed only on the output field.Inference uses vLLM with temperature 0.7, and reported timings exclude software initialization and checkpoint I/O overheads.Source: §G.3 · p. 24

G.4 EVALUATION

The evaluation covers four tool-use benchmarks and uses a soft-matching metric to assess functional correctness while discounting superficial string variations.

G.4 EVALUATION

  • The study evaluates NexusRaven, SealTool, API-Bank, and ToolAlpaca, spanning software operations, broad tool use, multi-turn API dialogue, and generalized tool use.The benchmarks contain 65, 4,076, 753, and 3,938 APIs or tool-use instances, respectively, as described in the passage.Source: §G.4 · p. 24
  • The test sets contain 318 NexusRaven, 294 SealTool, 361 API-Bank, and 103 ToolAlpaca samples, respectively.The API-Bank test set differs from the 316 Levels 1 and 2 samples identified in the benchmark description.Source: §G.4 · p. 24
  • Evaluation checks function names, arguments, and values or types, using soft matching to ignore case and minor morphological differences in string arguments.The authors report that this adjustment changes performance by only 2–3% and argue it better estimates functional correctness than exact matching or LLM-based judges.Source: §G.4 · p. 24

H ADDITIONAL RUN-TIME OVERHEAD AND RESOURCE USAGE ANALYSIS · I USE OF LLMS

On SealTool, TT-SI’s reported latency analysis shows faster processing than SFT despite using far fewer samples, while noting that excluded I/O operations can increase end-to-end latency. The paper also describes its uses of LLMs and safeguards for language refinement.

H ADDITIONAL RUN-TIME OVERHEAD AND RESOURCE USAGE ANALYSIS

  • The latency analysis excludes model merging and other I/O overheads, using HuggingFace for confidence estimation, vLLM for synthesis and inference, and LLaMA-Factory for training.The reported scope excludes merging and file I/O; the analysis measures TT-SI on SealTool.Source: H ADDITIONAL RUN-TIME OVERHEAD AND… · p. 24
  • Per-sample TT-SI steps average 0.87s for uncertainty estimation, 3.45s for synthesis, 2.05s for training updates, and 0.89s for inference.Together, the steps amount to an average of ∼7.3s per uncertain sample; the prose reports ∼36 minutes for 190 updates and 104 direct inferences.Source: H ADDITIONAL RUN-TIME OVERHEAD AND… · pp. 24–25
  • 7.26s for uncertain versus 1.76s for certain samples; TT-SI achieves a 3.7× wall-clock speed-up over SFT despite using ∼68× fewer samples.The table reports end-to-end average latency with merge and file I/O overhead excluded; the prose compares TT-SI with SFT training on SealTool’s 13K-sample split.Source: Table 4 · p. 25, H ADDITIONAL RUN-TIME OVERHEAD AND… · p. 25
  • Model merging, file-saving after training, and vLLM model loading can substantially increase end-to-end latency, so the authors recommend more optimized and scalable I/O strategies.They state that efficient file handling is outside their main contribution and note that third-party libraries offer ways to use merged weights without writing them to disk.Source: H ADDITIONAL RUN-TIME OVERHEAD AND… · p. 25

I USE OF LLMS

  • LLMs served as test-time training models, empirical baselines, and minor writing aids, with open-source Qwen and closed-source GPT-5-mini3 among the models used.The manuscript says both open- and closed-source models were employed for training and data generation.Source: I USE OF LLMS · p. 25
  • The authors state that all scientific ideas, methods, experiments, and conclusions are theirs, and that they reviewed language-refinement outputs to avoid hallucinated or incorrect content.The prompt for writing refinement asked for clearer sentences and removal of grammatical mistakes.Source: I USE OF LLMS · p. 25

Related papers

Loading 2510.07841v1…