CoT transformers simulate any Word RAM algorithm with poly-logarithmic overhead in three architectures, improving on quadratic TM overhead.
super hub Canonical reference
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
Canonical reference. 85% of citing Pith papers cite this work as background.
324
Pith papers citing it
Background
85% 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 48
dataset 3
method 3
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
UniQL is a human-verified benchmark providing aligned natural language questions and dialect-specific SQL queries for 16 SQL systems to evaluate cross-dialect generalization.
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.
Test-time training with KV binding reduces to learned linear attention.
Physics-IQ benchmark reveals that generative video models exhibit limited physical understanding unrelated to their visual quality.
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.
Variable-length autoregressive latent sequences, trained as variational inference with a PPO-style objective, give robot policies adaptive test-time compute and yield a reusable action tokenizer.
When reflections localize early errors, in-context search solves exp-small pass-rate problems with poly sequential attempts; otherwise it offers no asymptotic gain over parallel sampling, and the update is learnable and RLVR-optimal.
Multi-agent LLMs classify USPTO reactions and write verified SMIRKS rules, expanding a reaction taxonomy from 68 to 14,073 classes and matching proprietary classifiers on held-out and out-of-distribution data.
Preregistered placebo-controlled decomposition shows external executable counterexamples drive self-repair gains in small code models more than re-exposure or self-critique.
Empirical power-law frontier between predictive loss and structural forward work in LOB models extrapolates to held-out high-compute architectures with R²=0.941, motivating FastBiNLOB which exceeds SOTA macro-F1 at lower latency.
LBR performs token-level test-time scaling via local branch routing on hidden states, enabling end-to-end RL training and improving Pass@1 and Pass@32 on math benchmarks over CoT and RLVR baselines.
SPIRAL is a reinforcement learning framework that jointly optimizes sequential reasoning, parallel trace generation, and aggregation in language models for improved test-time performance.
SPOT-E uses entropy shaping on answer predictions with low-entropy anchors to optimize visual spotlights at test time via GRPO for better VLM performance on evidence-intensive tasks.
SWITCH uses explicit <swi> and </swi> boundary tokens to make latent chain-of-thought compatible with on-policy RL (GRPO) and open to causal mechanistic probing, outperforming prior hidden-state recurrence methods.
MARS is a margin-adversarial stopping rule for parallel LLM test-time scaling that saves 25-47% tokens while matching full-budget majority-vote accuracy by learning trace switch probabilities and applying adversarial bounds.
QGF performs test-time policy optimization for flow models in RL by guiding a behavior-cloned reference policy with value-function gradients, achieving strong results on high-dimensional offline RL benchmarks without additional policy training.
KCSAT-ML benchmark supplies human error rates for math problems and DRG metric exposes that model accuracy collapses on high-human-error items while test-time scaling shows non-monotonic gains and alignment failures.
PRISM is a contrastive, policy-aware training framework for process reward models that reduces false positives by 22% on PRMBench and boosts downstream accuracy up to 33% in Best-of-N selection by learning reliable relative comparisons instead of pointwise labels.
Three problem-level trajectory features derived from the distributional signature of failed LLM rollouts enable failure clustering at 84.3% accuracy and a training-free routing rule that improves rescue by 12.2% on hard cases.
TTT-RTL performs per-design test-time RL on an LLM policy with EDA-derived PPA rewards and an adaptive KL controller, reducing geometric-mean PPA product by 65.1% on RTLLM v2.0 and ADP by 59.4% on an industrial FPU unit.
LLMs achieve up to 78.8% accuracy and r=0.590 correlation mimicking individual SOEP respondents using cumulative microdata, with gains from more information but diminishing returns past the 75% entropy point.
Consequence-aware scheduler using an issue-text predictor routes more compute to high-cost failures and cuts cost-weighted loss by 22-33% versus difficulty-based allocation on SWE-bench tasks.
Rotate2Think estimates an orthogonal rotation from input to thinking embeddings via Procrustes analysis on a few examples and injects the resulting vector to prime reasoning traces, raising accuracy in 30 of 32 model-benchmark settings.
citing papers explorer
-
UniQL: Towards Dialect-Universal Benchmarking for Text-to-SQL
UniQL is a human-verified benchmark providing aligned natural language questions and dialect-specific SQL queries for 16 SQL systems to evaluate cross-dialect generalization.
-
When Does In-Context Search Help? A Sampling-Complexity Theory of Reflection-Driven Reasoning
When reflections localize early errors, in-context search solves exp-small pass-rate problems with poly sequential attempts; otherwise it offers no asymptotic gain over parallel sampling, and the update is learnable and RLVR-optimal.
-
Agentic generation of verifiable rules for deterministic, self-expanding reaction classification
Multi-agent LLMs classify USPTO reactions and write verified SMIRKS rules, expanding a reaction taxonomy from 68 to 14,073 classes and matching proprietary classifiers on held-out and out-of-distribution data.
-
SPIRAL: Learning to Search and Aggregate
SPIRAL is a reinforcement learning framework that jointly optimizes sequential reasoning, parallel trace generation, and aggregation in language models for improved test-time performance.
-
MARS: Margin-Adversarial Risk-controlled Stopping for Parallel LLM Test-time Scaling
MARS is a margin-adversarial stopping rule for parallel LLM test-time scaling that saves 25-47% tokens while matching full-budget majority-vote accuracy by learning trace switch probabilities and applying adversarial bounds.
-
Not All Errors Are Equal: Consequence-Aware Reasoning Compute Allocation
Consequence-aware scheduler using an issue-text predictor routes more compute to high-cost failures and cuts cost-weighted loss by 22-33% versus difficulty-based allocation on SWE-bench tasks.
-
The Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial Pressure
The paper identifies unfaithful capitulation, a failure mode where chain-of-thought remains correct but the emitted answer flips wrong under sustained adversarial pressure in multi-turn dialogue.
-
LaneRoPE: Positional Encoding for Collaborative Parallel Reasoning and Generation
LaneRoPE adds an inter-sequence attention mask and extended RoPE to enable collaborative parallel sequence generation in LLMs, yielding accuracy gains on math reasoning under length limits.
-
Co-ReAct: Rubrics as Step-Level Collaborators for ReAct Agents
A GRPO-trained rubric generator that injects step-level criteria into ReAct agents improves deep-research benchmark scores over several test-time compute baselines.
-
CAPS: Cascaded Adaptive Pairwise Selection for Efficient Parallel Reasoning
CAPS is a four-stage inference-only cascade that adapts how much of each solution the verifier sees and how comparisons are distributed, halving per-candidate verifier tokens while outperforming uniform pairwise verification on most benchmarks.
-
Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems
A survey that unifies prior work on multi-agent LLM systems via the LIFE framework, mapping dependencies across collaboration, failure attribution, and autonomous self-evolution while identifying cross-stage challenges.
-
Uncovering the Representation Geometry of Minimal Cores in Overcomplete Reasoning Traces
Language models produce overcomplete reasoning traces where on average 46% of steps can be removed while preserving the answer in 86% of cases, with necessity concentrated in the top three steps.
-
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.
-
Active Testing of Large Language Models via Approximate Neyman Allocation
Proposes surrogate semantic entropy stratification followed by approximate Neyman allocation for active testing of LLMs on generative benchmarks, reporting up to 28% MSE reduction and 22.9% average budget savings versus uniform sampling.
-
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.
-
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.
-
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.
-
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.
-
Small Generalizable Prompt Predictive Models Can Steer Efficient RL Post-Training of Large Reasoning Models
GPS trains a small model on optimization history to predict prompt difficulty and select intermediate-difficulty diverse batches, yielding better training efficiency, final performance, and test-time allocation than baselines on reasoning benchmarks.
-
Inference-Time Scaling of Verification: Self-Evolving Deep Research Agents via Test-Time Rubric-Guided Verification
DeepVerifier enables self-evolving deep research agents via rubric-guided verification at test time, delivering 8-11% accuracy gains on GAIA and XBench-DeepSearch subsets.
-
ToolPRM: Fine-Grained Inference Scaling of Structured Outputs for Function Calling
ToolPRM provides fine-grained intra-call process supervision via a new dataset and reward model, outperforming outcome and coarse-grained alternatives on function-calling benchmarks.
-
Bayesian Social Deduction with Graph-Informed Language Models
Hybrid Bayesian-graph LLM agent reaches competitive performance against large models and achieves 67% win rate against humans in controlled Avalon play, outperforming baselines and human teammates.
-
Recursive Self-Improvement in AI: From Bounded Self-Refinement to Autonomous Research Loops
A survey of 1,250 papers organizes AI self-improvement along two axes—what is improved and loop closure—finding that demonstrated self-improvement strength tracks a verification hierarchy from formal verifiers down to intrinsic self-assessment.
-
The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI
Swapping only the orchestration layer around six AI models cut token cost per task 41% and latency 44% with quality at parity, making orchestration a bigger cost lever than model choice.
-
Cost-Effective Agent Harnesses for Abstract Reasoning and Generalization on ARC-AGI-1
Explorer-Definer and Reflective Orchestrator harnesses raise DeepSeek V3.2 from 15.5% to 67.25% pass@2 on ARC-AGI-1 public eval at $0.25–$0.62 per task without ARC-specific training.
-
Active Inference as the Test-Time Scaling Law for Physical AI Agents
Active inference supplies a test-time scaling law for physical AI agents by framing policy updates as soft Bayesian inference that reduces expected prediction errors, with a variational solution shown to outperform Q-learning and Bayesian RL in driving simulations.
-
Calibration Is Not Control: Why LLM-Agent Oversight Needs Intervention
Action-conditioned estimation of intervention advantage via prefix branching reduces control regret over calibrated scalar risk scores in LLM agent oversight across benchmarks.
-
Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers
FPRM is a Transformer-based model using fixed-point convergence for adaptive halting in looped architectures, claimed effective on Sudoku, Maze, state-tracking, and ARC-AGI benchmarks.
-
LegalHalluLens: Typed Hallucination Auditing and Calibrated Multi-Agent Debate for Trustworthy Legal AI
LegalHalluLens provides typed hallucination profiles over CUAD, a Risk Direction Index, and a calibrated debate pipeline that reveals 38-40 pp category gaps hidden by aggregate 52% error rates and reduces fabricated detections by 45%.
-
How Inference Compute Shapes Frontier LLM Evaluation
Frontier LLM benchmark scores are protocol-dependent: larger token budgets, feedback, and repeated submissions raise scores substantially on some benchmarks (FrontierMath, HLE, cyber) but barely on others (HealthBench, SWE-Bench Pro).
-
GitOfThoughts: Version-Controlled Reasoning and Agent Memory You Can Replay, Diff, and Merge
GitOfThoughts stores agent reasoning as a git repo and shows memory from past problems improves accuracy only when new problems are nearly identical (cosine similarity >0.8), with self-consistency providing the main gain on novel tasks.
-
Momentum for Reasoning: Dense Intrinsic Signals in Policy Optimization
ISPO densifies GRPO rewards with sequence-level informativeness and token-level directional signals from policy probabilities to reduce zero-advantage collapse and hallucinated certainty on math benchmarks.
-
Scaffold Effects on GAIA: A Controlled Comparison
A controlled comparison shows scaffold choice alters GAIA Level 1-2 accuracy by up to 28 points, with effects varying by model family rather than capability tier alone.
-
When Does Delegation Beat Majority? A Delegation-Based Aggregator for Multi-Sample LLM Inference
Propagational Proxy Voting driven by letter entropy and centered reasoning embeddings beats majority by +2.24 pp on non-trivial MMLU-Pro questions without labels or training.
-
GITCO: Gated Inference-Time Context Optimization in TSFMs
GITCO delivers +1.95% average MASE reduction on TimesFM 2.5 across 53 datasets by gated inference-time suppression of anomalous patches, capturing 89.9% of the improvement upper bound.
-
Latent Reward Steering: An Adaptive Inference-Time Framework that Implicitly Promotes Cognitive Behaviors in Reasoning LLMs
Reward-guided optimization of SAE latents, gated by reward and confidence, improves multi-benchmark reasoning and is post-hoc associated with more verification and course correction.
-
ReasonOps: Operator Segmentation for LLM Reasoning Traces
Unsupervised clustering on sentence-initial 3-token pivots extracts 7 universal reasoning operators from 44k traces across 12 LLMs that enable model fingerprinting and answer-correctness prediction.
-
In LLM Reasoning, there is Irrationality on top of Value Misalignment
LLMs display widespread rational value risk in reasoning that value alignment reduces but does not remove, with risk sensitive to inference strategy and showing diminishing returns from longer reasoning.
-
OpenDeepThink: Parallel Reasoning via Bradley-Terry Aggregation
OpenDeepThink uses Bradley-Terry aggregation of LLM pairwise judgments to rank and evolve parallel reasoning traces, improving Gemini 3.1 Pro Codeforces Elo by 405 points over 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.
-
Engagement Process: Rethinking the Temporal Interface of Action and Observation
Engagement Process (EP) decouples actions and observations as independent event streams over time within a POMDP structure to explicitly model temporal dynamics in agent interactions.
-
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.
-
Distilling Long-CoT Reasoning through Collaborative Step-wise Multi-Teacher Decoding
CoRD uses collaborative multi-teacher step-wise decoding with perplexity-guided beam search to generate higher-quality Long-CoT data that lets smaller models reach near-teacher performance with less supervision.
-
Reasoning Structure Matters for Safety Alignment of Reasoning Models
Changing the internal reasoning structure of large reasoning models through simple supervised fine-tuning on 1K examples produces strong safety alignment that generalizes across tasks and languages.
-
When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
Extended reasoning in LLMs exhibits overthinking and diminishing returns, with optimal thinking length varying by problem difficulty, allowing significant compute savings by stopping at moderate budgets.
-
Conformal Thinking: Risk Control for Reasoning on a Compute Budget
Conformal risk control with upper and lower thresholds lets LLMs adaptively stop reasoning while guaranteeing a maximum error rate and minimizing token use.
-
MobiBench: Multi-Branch, Modular Benchmark for Mobile GUI Agents
MobiBench reaches near-human offline evaluation fidelity for mobile GUI agents by accepting any valid action at each step, and enables modular attribution of performance to agent components.
-
Glia: A Human-Inspired AI for Automated Systems Design and Optimization
Glia deploys a multi-agent LLM workflow with reasoning, experimentation, and analysis agents to generate interpretable algorithms for request routing, scheduling, and auto-scaling in distributed GPU clusters, reaching human-expert performance levels.
-
DeepSearch: Overcome the Bottleneck of Reinforcement Learning with Verifiable Rewards via Monte Carlo Tree Search
DeepSearch embeds MCTS into RLVR training with global frontier selection, entropy guidance, and adaptive replay to achieve 62.95% average accuracy on math reasoning benchmarks while using 5.7x fewer GPU hours than extended training.