AutoTTS discovers width-depth test-time scaling controllers through agentic search in a pre-collected trajectory environment, yielding better accuracy-cost tradeoffs than hand-designed baselines on math reasoning tasks at low cost.
super hub Mixed citations
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
Mixed citation behavior. Most common role is background (40%).
127
Pith papers citing it
Background
40% of classified citations
abstract
Enabling LLMs to improve their outputs by using more test-time computation is a critical step towards building generally self-improving agents that can operate on open-ended natural language. In this paper, we study the scaling of inference-time computation in LLMs, with a focus on answering the question: if an LLM is allowed to use a fixed but non-trivial amount of inference-time compute, how much can it improve its performance on a challenging prompt? Answering this question has implications not only on the achievable performance of LLMs, but also on the future of LLM pretraining and how one should tradeoff inference-time and pre-training compute. Despite its importance, little research attempted to understand the scaling behaviors of various test-time inference methods. Moreover, current work largely provides negative results for a number of these strategies. In this work, we analyze two primary mechanisms to scale test-time computation: (1) searching against dense, process-based verifier reward models; and (2) updating the model's distribution over a response adaptively, given the prompt at test time. We find that in both cases, the effectiveness of different approaches to scaling test-time compute critically varies depending on the difficulty of the prompt. This observation motivates applying a "compute-optimal" scaling strategy, which acts to most effectively allocate test-time compute adaptively per prompt. Using this compute-optimal strategy, we can improve the efficiency of test-time compute scaling by more than 4x compared to a best-of-N baseline. Additionally, in a FLOPs-matched evaluation, we find that on problems where a smaller base model attains somewhat non-trivial success rates, test-time compute can be used to outperform a 14x larger model.
hub tools
citation-role summary
background 3
dataset 2
citation-polarity summary
claims ledger
- abstract Enabling LLMs to improve their outputs by using more test-time computation is a critical step towards building generally self-improving agents that can operate on open-ended natural language. In this paper, we study the scaling of inference-time computation in LLMs, with a focus on answering the question: if an LLM is allowed to use a fixed but non-trivial amount of inference-time compute, how much can it improve its performance on a challenging prompt? Answering this question has implications not only on the achievable performance of LLMs, but also on the future of LLM pretraining and how one
authors
co-cited works
representative citing papers
MedPRMBench is the first fine-grained benchmark for process reward models in medical reasoning, featuring 6500 questions, 13000 chains, 113910 step labels, and a baseline that improves downstream QA accuracy by 3.2-6.7 points.
EvoLib enables LLMs to accumulate, reuse, and evolve knowledge abstractions from inference trajectories at test time, yielding substantial gains on math reasoning, code generation, and agentic benchmarks without parameter updates or supervision.
QueST adapts LLMs at test time by generating query-specific problem-solution pairs for self-supervised fine-tuning, improving reasoning performance without external data.
MOPD improves on-policy distillation for LLMs by using peer successes for positive patterns and failures for negative examples to create more informative teacher signals.
VeGAS improves MLLM-based embodied agents by sampling action ensembles and using a verifier trained on LLM-synthesized failure cases, yielding up to 36% relative gains on hard multi-object long-horizon tasks in Habitat and ALFRED.
StepCodeReasoner aligns code reasoning with verifiable stepwise execution traces via print anchors and bi-level GRPO reinforcement learning, reaching SOTA results on CRUXEval (91.1%) and LiveCodeBench (86.5%) for a 7B model.
A softmax-weighted centroid of the local top-K documents interpolated with the query improves nDCG@10 for frozen embedding models across seven families on held-out BEIR data.
V-ABS is an action-observer beam search method with entropy-based adaptive weighting and an 80k-sample SFT dataset that delivers 19.7% average gains on visual reasoning tasks for MLLMs.
RubricRefine improves tool-use agent reliability to 0.86 on M3ToolEval by generating rubrics for pre-execution contract checking and iterative repair, outperforming baselines at 2.6X lower latency while showing no gain on single-step tasks.
BubbleSpec exploits long-tail bubbles in synchronous RL by using faster ranks' idle time to pre-generate rollout drafts for speculative decoding, reducing steps by 50% and raising throughput up to 1.8x while preserving exact synchrony.
CollabVR improves video reasoning performance by coupling vision-language models and video generation models in a closed-loop step-level collaboration that detects and repairs generation failures.
DUET improves RLVR by allocating tokens across both prompt selection and rollout length, outperforming full-budget baselines even when using only half the tokens.
CA-SQL achieves 51.72% execution accuracy on the challenging tier of the BIRD benchmark using GPT-4o-mini by scaling exploration breadth according to estimated task difficulty, evolutionary prompt seeding, and candidate voting.
POPO uses bounded importance sampling on positive rollouts and a siamese policy network to achieve implicit negative gradients and stable optimization, matching or exceeding GRPO on math benchmarks such as 36.67% on AIME 2025.
Joint Consistency casts test-time aggregation as Ising-type energy minimization with pairwise LLM-judge interactions, subsuming voting methods and outperforming baselines across reasoning tasks.
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
LOVER creates an unsupervised logic-regularized verifier that reaches 95% of supervised verifier performance on reasoning tasks across 10 datasets.
BLADE uses Bayesian list-wise alignment with dynamic estimation to create a self-evolving target that overcomes limitations of static references in LLM-based recommendation, yielding sustained gains in ranking and complex metrics.
POSTCONDBENCH is a new multilingual benchmark that evaluates LLM postcondition generation on real code using defect discrimination to assess completeness beyond surface matching.
LiveFMBench shows that direct LLM prompting for C program formal specs overestimates accuracy by ~20% due to unfaithful behaviors like deceiving provers, while agentic workflows help under low sampling but overall performance remains far below human-authored specs.
TokenArena is a continuous benchmark for AI inference endpoints that measures output speed, time to first token, blended price, effective context, quality, and modeled energy to produce composites of joules per correct answer, dollars per correct answer, and endpoint fidelity.
In a controlled arithmetic-grammar program synthesis environment, diverse sampling across semantic and syntactic spaces yields robust density generalization while support generalization for novel syntax remains poor, with performance falling over 30 percent and compute scaling following a strictly 1
Semantic-level and verification-based uncertainty methods outperform token-level baselines for audio reasoning in ALLMs, but their relative performance on hallucination and unanswerable-question benchmarks is model- and task-dependent.
citing papers explorer
-
LLMs Improving LLMs: Agentic Discovery for Test-Time Scaling
AutoTTS discovers width-depth test-time scaling controllers through agentic search in a pre-collected trajectory environment, yielding better accuracy-cost tradeoffs than hand-designed baselines on math reasoning tasks at low cost.
-
MedPRMBench: A Fine-grained Benchmark for Process Reward Models in Medical Reasoning
MedPRMBench is the first fine-grained benchmark for process reward models in medical reasoning, featuring 6500 questions, 13000 chains, 113910 step labels, and a baseline that improves downstream QA accuracy by 3.2-6.7 points.
-
Test-Time Learning with an Evolving Library
EvoLib enables LLMs to accumulate, reuse, and evolve knowledge abstractions from inference trajectories at test time, yielding substantial gains on math reasoning, code generation, and agentic benchmarks without parameter updates or supervision.
-
Query-Conditioned Test-Time Self-Training for Large Language Models
QueST adapts LLMs at test time by generating query-specific problem-solution pairs for self-supervised fine-tuning, improving reasoning performance without external data.
-
Multi-Rollout On-Policy Distillation via Peer Successes and Failures
MOPD improves on-policy distillation for LLMs by using peer successes for positive patterns and failures for negative examples to create more informative teacher signals.
-
Think Twice, Act Once: Verifier-Guided Action Selection For Embodied Agents
VeGAS improves MLLM-based embodied agents by sampling action ensembles and using a verifier trained on LLM-synthesized failure cases, yielding up to 36% relative gains on hard multi-object long-horizon tasks in Habitat and ALFRED.
-
StepCodeReasoner: Aligning Code Reasoning with Stepwise Execution Traces via Reinforcement Learning
StepCodeReasoner aligns code reasoning with verifiable stepwise execution traces via print anchors and bi-level GRPO reinforcement learning, reaching SOTA results on CRUXEval (91.1%) and LiveCodeBench (86.5%) for a 7B model.
-
Test-Time Compute for Dense Retrieval: Agentic Program Generation with Frozen Embedding Models
A softmax-weighted centroid of the local top-K documents interpolated with the query improves nDCG@10 for frozen embedding models across seven families on held-out BEIR data.
-
V-ABS: Action-Observer Driven Beam Search for Dynamic Visual Reasoning
V-ABS is an action-observer beam search method with entropy-based adaptive weighting and an 80k-sample SFT dataset that delivers 19.7% average gains on visual reasoning tasks for MLLMs.
-
RubricRefine: Improving Tool-Use Agent Reliability with Training-Free Pre-Execution Refinement
RubricRefine improves tool-use agent reliability to 0.86 on M3ToolEval by generating rubrics for pre-execution contract checking and iterative repair, outperforming baselines at 2.6X lower latency while showing no gain on single-step tasks.
-
BubbleSpec: Turning Long-Tail Bubbles into Speculative Rollout Drafts for Synchronous Reinforcement Learning
BubbleSpec exploits long-tail bubbles in synchronous RL by using faster ranks' idle time to pre-generate rollout drafts for speculative decoding, reducing steps by 50% and raising throughput up to 1.8x while preserving exact synchrony.
-
CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models
CollabVR improves video reasoning performance by coupling vision-language models and video generation models in a closed-loop step-level collaboration that detects and repairs generation failures.
-
DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards
DUET improves RLVR by allocating tokens across both prompt selection and rollout length, outperforming full-budget baselines even when using only half the tokens.
-
CA-SQL: Complexity-Aware Inference Time Reasoning for Text-to-SQL via Exploration and Compute Budget Allocation
CA-SQL achieves 51.72% execution accuracy on the challenging tier of the BIRD benchmark using GPT-4o-mini by scaling exploration breadth according to estimated task difficulty, evolutionary prompt seeding, and candidate voting.
-
Beyond Negative Rollouts: Positive-Only Policy Optimization with Implicit Negative Gradients
POPO uses bounded importance sampling on positive rollouts and a siamese policy network to achieve implicit negative gradients and stable optimization, matching or exceeding GRPO on math benchmarks such as 36.67% on AIME 2025.
-
Joint Consistency: A Unified Test-Time Aggregation Framework via Energy Minimization
Joint Consistency casts test-time aggregation as Ising-type energy minimization with pairwise LLM-judge interactions, subsuming voting methods and outperforming baselines across reasoning tasks.
-
Post Reasoning: Improving the Performance of Non-Thinking Models at No Cost
Post-Reasoning boosts LLM accuracy by reversing the usual answer-after-reasoning order, delivering mean relative gains of 17.37% across 117 model-benchmark pairs with zero extra cost.
-
Logic-Regularized Verifier Elicits Reasoning from LLMs
LOVER creates an unsupervised logic-regularized verifier that reaches 95% of supervised verifier performance on reasoning tasks across 10 datasets.
-
Beyond Static Best-of-N: Bayesian List-wise Alignment for LLM-based Recommendation
BLADE uses Bayesian list-wise alignment with dynamic estimation to create a self-evolving target that overcomes limitations of static references in LLM-based recommendation, yielding sustained gains in ranking and complex metrics.
-
POSTCONDBENCH: Benchmarking Correctness and Completeness in Formal Postcondition Inference
POSTCONDBENCH is a new multilingual benchmark that evaluates LLM postcondition generation on real code using defect discrimination to assess completeness beyond surface matching.
-
LiveFMBench: Unveiling the Power and Limits of Agentic Workflows in Specification Generation
LiveFMBench shows that direct LLM prompting for C program formal specs overestimates accuracy by ~20% due to unfaithful behaviors like deceiving provers, while agentic workflows help under low sampling but overall performance remains far below human-authored specs.
-
Token Arena: A Continuous Benchmark Unifying Energy and Cognition in AI Inference
TokenArena is a continuous benchmark for AI inference endpoints that measures output speed, time to first token, blended price, effective context, quality, and modeled energy to produce composites of joules per correct answer, dollars per correct answer, and endpoint fidelity.
-
Beyond the Training Distribution: Mapping Generalization Boundaries in Neural Program Synthesis
In a controlled arithmetic-grammar program synthesis environment, diverse sampling across semantic and syntactic spaces yields robust density generalization while support generalization for novel syntax remains poor, with performance falling over 30 percent and compute scaling following a strictly 1
-
Walking Through Uncertainty: An Empirical Study of Uncertainty Estimation for Audio-Aware Large Language Models
Semantic-level and verification-based uncertainty methods outperform token-level baselines for audio reasoning in ALLMs, but their relative performance on hallucination and unanswerable-question benchmarks is model- and task-dependent.
-
Self-Correction as Feedback Control: Error Dynamics, Stability Thresholds, and Prompt Interventions in LLMs
Self-correction in LLMs is stable and non-degrading only when ECR/EIR exceeds initial accuracy over (1-accuracy), with EIR below 0.5% cleanly separating helpful from harmful cases across models.
-
PAC-MCTS: Bias-Aware Pruning for Robust LLM-Guided Search and Planning
PAC-MCTS supplies bias-aware confidence bounds for pruning in LLM-guided MCTS, with O((Δ-4L)^{-2}) upper and Ω((Δ-2L)^{-2}) lower sample-complexity guarantees and up to 78% fewer API calls on Blocksworld and ALFWorld.
-
Towards Unconstrained Human-Object Interaction
Introduces the U-HOI task and shows MLLMs plus a language-to-graph pipeline can handle human-object interactions without any predefined vocabulary at training or inference time.
-
AdverMCTS: Combating Pseudo-Correctness in Code Generation via Adversarial Monte Carlo Tree Search
AdverMCTS frames code generation as a minimax game where an attacker evolves tests to expose flaws in solver-generated code, yielding more robust outputs than static-test baselines.
-
AI Achieves a Perfect LSAT Score
Language models achieve a perfect LSAT score, with experiments showing that internal thinking phases and a fine-tuned process reward model are key to high performance on logical reasoning questions.
-
Can LLMs Deobfuscate Binary Code? A Systematic Analysis of Large Language Models into Pseudocode Deobfuscation
LLM deobfuscation of binaries to pseudocode depends more on reasoning ability and task-specific fine-tuning than on model size, with reasoning models showing robustness across ISAs and obfuscation levels on the new BinDeObfBench.
-
Sampling for Quality: Training-Free Reward-Guided LLM Decoding via Sequential Monte Carlo
Sequential Monte Carlo sampling from a reward-augmented sequence distribution improves LLM performance on HumanEval by up to 54.9% and MATH500 by up to 8.8%, outperforming standard sampling and GRPO.
-
Less is More: Recursive Reasoning with Tiny Networks
TRM with 7M parameters achieves 45% accuracy on ARC-AGI-1 and 8% on ARC-AGI-2, surpassing most LLMs with under 0.01% of their parameters.
-
Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs
o1-like models overthink easy tasks; self-training reduces compute use without accuracy loss on GSM8K, MATH500, GPQA, and AIME.
-
Training Large Language Models to Reason in a Continuous Latent Space
Coconut lets LLMs perform reasoning directly in continuous latent space by recycling hidden states as inputs, outperforming standard chain-of-thought on search-intensive logical tasks with better accuracy-efficiency trade-offs.
-
Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis
DataPRM is a new process reward model for data analysis agents that detects silent errors via environment interaction and ternary rewards, yielding 7-11% gains on benchmarks and further RL improvements.
-
Efficient Test-Time Inference via Deterministic Exploration of Truncated Decoding Trees
Distinct Leaf Enumeration (DLE) replaces stochastic self-consistency sampling with deterministic traversal of a truncated decoding tree to enumerate distinct leaves, increasing coverage and reducing redundant computation while improving performance on math, coding, and reasoning benchmarks.
-
Self-Consistency from Only Two Samples: CoT-PoT Ensembling for Efficient LLM Reasoning
CoT-PoT ensembling achieves self-consistency accuracy in LLMs with only two samples for 78.6% of tasks, reducing computation by 9.3x compared to standard methods.
-
Does RL Expand the Capability Boundary of LLM Agents? A PASS@(k,T) Analysis
RL expands the capability boundary of LLM agents on compositional tool-use tasks, shown by non-converging pass curves at large k with increasing T, while SFT regresses it and the effect is absent on simpler tasks.
-
OpenDeepThink: Parallel Reasoning via Bradley--Terry Aggregation
OpenDeepThink improves LLM reasoning by ranking parallel candidate traces via Bradley-Terry aggregation of LLM pairwise judgments, achieving a +405 Codeforces Elo gain on Gemini 3.1 Pro after eight rounds.
-
Nice Fold or Hero Call: Learning Budget-Efficient Thinking for Adaptive Reasoning
BET reduces reasoning tokens by about 55% on average while improving performance across benchmarks by learning to short-solve easy queries, fold early on unsolvable ones, and preserve budget for hard solvable queries.
-
Decaf: Improving Neural Decompilation with Automatic Feedback and Search
Decaf uses compiler feedback and search to improve neural decompilation, boosting semantic success rate from 26.0% to 83.9% on ExeBench Real -O2 split.
-
Engagement Process: Rethinking the Temporal Interface of Action and Observation
Engagement Process decouples actions and observations into separate time-based event streams within a POMDP structure to explicitly model timing mismatches, deliberation latency, and multi-rate interactions.
-
fg-expo: Frontier-guided exploration-prioritized policy optimization via adaptive kl and gaussian curriculum
FG-ExPO improves GRPO by adaptively scaling the KL penalty with batch accuracy and sampling questions via a Gaussian centered at 0.5 accuracy, delivering up to 13.34 point gains on AIME 2025 pass@32.
-
Likelihood scoring for continuations of mathematical text: a self-supervised benchmark with tests for shortcut vulnerabilities
A new benchmark uses separate predictor and scorer LLMs to test whether forecast strings improve likelihood of hidden mathematical equation continuations, with controls that detect priming shortcuts.
-
What should post-training optimize? A test-time scaling law perspective
Tail-extrapolated estimators approximate best-of-N policy gradients from limited training rollouts by leveraging upper-tail reward statistics under structural assumptions.
-
Breaking the Reward Barrier: Accelerating Tree-of-Thought Reasoning via Speculative Exploration
SPEX delivers 1.2-3x speedup on ToT algorithms via speculative path selection, dynamic budget allocation, and adaptive early termination, reaching up to 4.1x when combined with token-level speculative decoding.
-
Active Testing of Large Language Models via Approximate Neyman Allocation
Active testing via surrogate semantic entropy stratification and approximate Neyman allocation reduces MSE by up to 28% versus uniform sampling and saves about 23% of the labeling budget on language and multimodal benchmarks.
-
Forge: Quality-Aware Reinforcement Learning for NP-Hard Optimization in LLMs
OPT-BENCH trains LLMs on NP-hard optimization via quality-aware RLVR, achieving 93.1% success rate and 46.6% quality ratio on Qwen2.5-7B while outperforming GPT-4o and transferring gains to other domains.
-
Hint Tuning: Less Data Makes Better Reasoners
Hint Tuning uses an instruct model as a difficulty probe to create 1K multi-level hint examples that train reasoning models to calibrate chain-of-thought length, cutting tokens by 31.5% on average across 4B-32B models without accuracy loss.
-
AIPO: : Learning to Reason from Active Interaction
AIPO trains LLMs to expand their reasoning capability boundary via active multi-agent interaction with Verify, Knowledge, and Reasoning agents during RLVR, using importance sampling and clipping to handle feedback, then drops the agents at inference.