LatentEvolve: Self-Evolving Test-Time Scaling in Latent Space

• Test-time scaling aims to harness LLMs’ intrinsic knowledge without additional training during inference, as training-time scaling faces resource demands and limits in high-quality data.The paper notes that pre-training scaling has begun to slow.Source: §1 · p. 1
• Most test-time scaling methods treat distinct queries as independent, so prior successes do not refine later selection strategies or transfer reflective strategies to new tasks.The paper frames learning from experience while solving more problems as the central design question.Source: §1 · p. 2
• LatentEvolve addresses inter-task independence with a dual-phase framework inspired by complementary learning systems theory.Daytime scaling optimizes each query’s latent state using retrieved traces from solved problems; nighttime scaling trains a compact latent weaver on daytime traces to generate initial latent representations for future tasks.Source: §1 · p. 2
• The daytime and nighttime processes alternate to support fast and slow evolution without ground-truth labels or other external signals.The daytime phase accumulates experiential knowledge, which nighttime reflection periodically consolidates into procedural memory.Source: §1 · p. 2
• Up to 23.3% gains on math reasoning; LatentEvolve also surpasses GRPO and LatentSeek on MATH-500 by 1.75% and 11.40%, respectively, and transfers gains to GPQA and JAMA.The reported out-of-domain gains are 7.07% on GPQA and 5.22% on JAMA; scaling across new domains does not degrade previously seen domains and can modestly improve them.Source: §1 · p. 2

• Test-time training updates model parameters during testing, whereas test-time scaling increases computation without altering them.TTS can operate in explicit natural-language space or latent space.Source: §2 · p. 3
• Latent computation methods either enable native latent reasoning or use latent representations to steer generation; LatentEvolve adds dual-stage test-time evolution in latent space.Prior latent strategies are described as generally remaining inter-task independent.Source: §2 · p. 3
• LLM evolution approaches use parameter updates, experience databases, or skill condensation, while LatentEvolve treats latent sequences as a compact, adaptable skill repository.Source: §2 · p. 3

• Latent-space aided reasoning conditions a frozen language model’s output generation on an auxiliary sequence of continuous vectors called latent tokens.The latent sequence acts as an instance-specific control signal, and the output is sampled conditioned on both context and latent intervention.Source: §3 · p. 3
• The latent intervention is selected to maximize an objective measuring the expected quality of generated outputs.The quality score Q(y) evaluates an output’s quality.Source: §3 · p. 3
• Existing approaches either optimize shared task-specific soft prompts or query-specific latent paths, but most do not rapidly learn or evolve across instances.The paper identifies this lack of on-the-fly adaptation from cumulative experience as a limitation of those practices.Source: §3 · p. 4

4 METHODOLOGY

• The methodology pairs fast, instance-specific adaptation with slower consolidation of accumulated experiences into a compact parametric prior.The two phases form a recurrent cycle of latent-space evolution.Source: §4 · p. 4

4.1 DAYTIME TEST-TIME OPTIMIZATION

• For each query, LatentEvolve retrieves relevant experiences, initializes from a CoT latent prefix, transfers weighted optimization momentum, then refines and archives high-quality results.Momentum is the change from an experience’s initial to refined latent state; refinement uses self-assigned quality scores and stops after K iterations or three successive rounds without an increase.Source: §4.1 · pp. 4–5

4.2 NIGHTTIME DELIBERATIVE CONSOLIDATION

• Periodically, the latent weaver trains on accumulated experience triplets to consolidate episodic optimizations into procedural knowledge for future reasoning.The weaver is trained to predict refined latent sequences from context embeddings and base states.Source: §4.2 · pp. 5–6

4.3 DUAL-PHASE EVOLVING SCALING

• LatentEvolve cycles between weaver-based initialization, daytime optimization, and nighttime consolidation, refreshing episodic traces while building generalized procedural knowledge without external labels.Nighttime scaling is invoked every 200 test-time instances.Source: §4.3 · p. 6

5.1 EXPERIMENT SETTING

• Experiments span eight benchmarks across general QA, mathematical, scientific, and medical reasoning, using five LLM backbones and Pass@1 accuracy at temperature 0.AIME24/25 are evaluated after applying LatentEvolve on MATH-500; the latent weaver is Qwen-2.5-1.5b.Source: §5.1 · p. 6

5.2 MAIN RESULTS

• LatentEvolve matches or surpasses TTRL on general QA and improves other domains, including MMLU +6.4% and SciBench +8.52% with Qwen2.5-7b.The comparisons use two backbones and eight benchmarks against thirteen baselines.Source: §5.2 · p. 7, Table 1 · p. 7
• LatentEvolve consistently improves results across five backbones, with MATH-500 gains increasing from 9.6% on Llama3.2-3b to 20.8% on Gemma-3-12b.Table 2 compares vanilla models with LatentEvolve-enhanced models and reports absolute improvements in its Δ row.Source: §5.2 · p. 8, Table 2 · p. 8

5.3 GENERALIZATION AND CONTINUAL LEARNING STUDY

• After two MATH rounds, performance rises from 57.6% to 78.6% in-domain, with JAMA +6.6% and MMLU +1.5%; nighttime scaling yields a larger first-round JAMA gain than daytime scaling.After two MMLU rounds, MATH improves from 78.6% to 80.2%. Figure 3 tracks Gemma-3-12b across in-domain MATH and MMLU and out-of-domain GPQA and JAMA.Source: §5.3 · p. 8, Figure 3 · p. 8

5.4 FRAMEWORK ANALYASIS

• Performance peaks at L′ = 30 as its value increases from 10 to 50, reaching MATH 80.3% and SciBench 33.4%.The authors suggest too few dimensions may encode experience inadequately, while too many may hinder daytime self-supervised refinement.Source: §5.4 · p. 9, Figure 4 · p. 9
• Removing either component reduces performance; on SciBench at L′ = 30, removing daytime scaling drops accuracy from 33.4% to 28.5%, while removing nighttime scaling causes a 6.8% decline.Figure 4 presents the ablation study and parameter sensitivity analysis.Source: §5.4 · p. 9, Figure 4 · p. 9
• LatentEvolve generations include unusual reasoning tokens and patterns but still produce correct answers, and their reasoning uses fewer decoding tokens than vanilla CoT.Table 3 presents questions from SciBench and GPQA, answers, and LatentEvolve generations, with peculiar reasoning tokens highlighted.Source: §5.4 · p. 9, Table 3 · p. 9

6 CONCLUSION

• Across eight benchmarks and five model backbones, LatentEvolve surpasses state-of-the-art TTS methods, transfers effectively across tasks, and exhibits steady continual learning.The paper frames this as a path toward test-time scaling that also evolves through inference.Source: §6 · p. 10
• Alternating daytime scaling for fast episodic adaptation with nighttime scaling for slow procedural consolidation lets LLMs accumulate and refine experiential knowledge during inference without external supervision.Source: §6 · p. 10

A.1 SELF-SUPERVISED LATENT REFINEMENT

• The self-rewarding function aggregates LLM-assigned scores for answer correctness, comprehension, calculation validity, format conformity, and clarity.Source: §A.1 · p. 17
• The weighted aggregation assigns weights 1:1:1:2:2, emphasizing format conformity and clarity over the other criteria.Source: §A.1 · p. 17
• The correctness prompt checks the final answer against the task and reasoning, using only internal consistency and accuracy rather than external knowledge.Source: §A.1 · p. 17
• The comprehension and calculation prompts assess coverage of task constraints and validity of numerical calculations and logical steps.Source: §A.1 · p. 17
• The format and clarity prompts assess compliance with requested answer presentation and whether the solution is readable, explicit, and complete.Source: §A.1 · p. 18

A.2 NIGHTTIME CONSOLIDATION

• Only latent triplets associated with sufficient confidence are retained for the latent weaver’s training, since its initial reasoning seeds should come from relatively high-quality data.Source: §A.2 · p. 18
• Confidence is defined by expected final-output quality after iterative refinement, and the admission threshold τ is set to 0.5 across experiments.Source: §A.2 · pp. 18–19

A.3 EVALUATION

• The prompt summary covers numerical-answer datasets including GSM8K, MATH, AIME 2024/2025, and SciBench, as well as multiple-choice datasets including MMLU, SciBench, JAMA, and GPQA.Source: §A.3 · p. 19
• SciBench appears in both listed prompt categories: numerical-answer and multiple-choice datasets.Source: §A.3 · p. 19
• Table 4 summarizes evaluation prompts used by LatentEvolve and other baselines.Source: Table 4 · p. 19, §A.3 · p. 19

A.4 DATASET DETAILS

• The evaluation uses 1,000 MMLU instances, MATH-500, all 1,319 GSM8K test problems, and 198 GPQA-Diamond questions.The GPQA-Diamond subset contains graduate-level questions of elevated difficulty.Source: §A.4 · p. 19
• SciBench contributes all 692 tasks, JAMA Clinical Challenge contributes all 1,511 test items, and AIME 2024 and 2025 each contribute 30 problems.Source: §A.4 · p. 19

A.6 SENSITIVITY ANALYSIS

• Table 5 reports LatentEvolve-to-vanilla-CoT answer-length ratios for full LatentEvolve, without Daytime, and without Nighttime.Source: Table 5 · p. 20
• At T=200, MMLU accuracy rises from 68.3% at Iter 1 to 73.6% at Iter 5; gains also appear at T=300 and T=500 but diminish as the interval grows.At T=1000, evolution is effectively reduced to one daytime/nighttime interaction, limiting gradual refinement; moderate intervals balance stability and adaptability.Source: §A.6 · p. 19, Table 6 · p. 20