arXiv subjects

Song Han

Publications and source records attributed to Song Han.

At least 19 recordsLinked to original sources

arXiv2026

Weaken to Strengthen: Weakening-Induced On-Policy Self-Distillation for Large Reasoning Models

Self-distillation offers a way to improve reasoning models when a suitable external teacher is unavailable. However, common approaches rely on privileged information to construct a stronger self-teacher, limiting their use when only unlabeled prompts are available. In this paper, we introduce Weakening-Induced On-Policy Self-Distillation (WISD), which instead constructs a weaker counterpart of the original model and uses the contrast between the two as a training signal without requiring privileged information. We explore two ways to construct the weaker model: low-bit quantization and magnitude pruning. We evaluate WISD across diverse architectures and scales on mathematical and coding reasoning tasks. Across these benchmarks, quantization and pruning improve average accuracy over the original models by 4.0 and 5.5 percentage points, respectively, achieving state-of-the-art performance even without privileged information. Notably, pruning raises Qwen3.5-4B's accuracy from 57.7% to 70.5% on HMMT 2025 and from 39.1% to 56.7% on LiveCodeBench v6. Quantization raises Qwen3.6-35B-A3B's accuracy from 68.5% to 74.6% and from 61.8% to 69.5% on the same benchmarks, respectively. These results establish WISD as a novel and effective self-distillation framework for large reasoning models.

cs.AI↗

arXiv2026

Delta-Matching: Closing the Final Gap of Native 8-bit Training for LLMs

Reliable FP8 attention remains a barrier to fully native 8-bit large language model training. We derive how forward-backward inconsistencies produce stale delta and empirically show how it distorts training dynamics. Our stale-delta hybrid runs show a modest loss gap at 569M parameters but substantial loss increases and downstream degradation at 1.67B and 5.29B. QK normalization, NoPE (no positional encoding), and lower-learning-rate context extension mitigate or delay degradation without eliminating it. This pattern suggests accumulated optimization error that smaller models and short runs can conceal. We propose Delta-Matching, proving that it restores the softmax gradient's zero-row-sum invariant under the stated numerical assumptions. It enables native block-scaled FP8 in every forward and backward attention-core matmul without architectural changes, smaller global batches, or auxiliary forward outputs. Across tested architectures, scales, and training stages, Delta-Matching matches BF16/FP32 mixed-precision training loss and overall downstream performance. We will release our implementation, trained models, and data recipes (https://jet-ai-projects.github.io/delta-matching).

cs.LG↗

arXiv2026

Lightning Weave: Improving the Accuracy-Efficiency Frontier of Reasoning Models through Capability Composition

A core goal of efficient reasoning is to improve the accuracy-efficiency frontier. However, jointly improving reasoning accuracy and inference efficiency can be challenging, as the two objectives can favor different reasoning behaviors. Independently post-trained models already offer distinct strengths in accuracy and efficiency. We introduce Lightning Weave, a post-training framework that extracts and composes these independently learned capabilities in a single student through on-policy distillation. Each acquired capability is represented by the policy shift from the model before post-training to the resulting specialist. Lightning Weave combines aligned log-ratio shifts at shared student token states and uses Tilted-Target DOPD to convert the cached signals into a stable learning target. Each anchor pair scores the cached trajectories once, enabling subsequent student training without serving multiple live anchor models concurrently. Across diverse student models and benchmarks in mathematics and code, Lightning Weave substantially improves upon the base students and achieves a state-of-the-art accuracy-efficiency frontier. On Qwen3.5-4B, it raises HMMT 2025 accuracy from 59.2% to 64.0% with 10.7% fewer response tokens, and LiveCodeBench v5 accuracy from 41.7% to 54.2% with 9.6% fewer response tokens. Adjusting the relative strengths of the anchor signals yields a strong empirical accuracy-efficiency Pareto frontier. These results establish Lightning Weave as a new practical route to efficient reasoning through capability composition. Code is released at https://github.com/jet-ai-projects/Lightning-Weave.

cs.AI↗

arXiv2026

Robota-Embed: Unified VLA Embeddings for Robot In-Context Learning and Evolution

Retrieving useful context for a robot requires matching its task, execution stage, and the continuation that should follow. We introduce Robota-Embed, to our knowledge the first vision-language-action (VLA) embedding model to jointly encode all three modalities in both current-state queries and recorded future continuations. Predictive retrieval training aligns these endpoints, enabling historical continuations to serve as context at inference time. We construct a multi-source corpus of 1,695,684 cleaned training anchors and introduce Robota-Embed-Bench, with 4,000 queries and a shared 8,000-entry gallery. Robota-Embed achieves 47.33% clip-level and 81.77% episode-level R@1, exceeding action-free Gemini Embedding 2 by 22.09 and 33.47 percentage points. Controlled retrieval studies show that removing native action degrades retrieval more than removing vision. We study two complementary applications. Embed2ICL uses retrieved experience for training-free and training-based policy conditioning, increasing success from 47.0% to 56.3% in DROID-Sim and from 86.4% to 92.8% on the LIBERO difficult subset, while improving task progress on real robots. Embed2Evolve uses execution feedback to grow task-local memory and refine context use around a frozen GPT-6 Astra policy. Across five ManiSkill tasks, it achieves 68% success, compared with 24% for GPT-6 Astra Direct and 33% for a harness independently evolved with Omni-Embed-Nemotron-3B under the same protocol. Relative to the latter, it uses approximately 79.5% fewer policy tokens per test rollout, counting both successes and failures. These results show that a unified, action-aware retrieval space can support robot in-context learning and task-local context evolution.

cs.RO↗

arXiv2026

Parason: Revealing Subtask and Trial Parallelism in LLM Reasoning

Scaling test-time reasoning has substantially improved the problem-solving ability of large language models (LLMs), but standard autoregressive decoding still executes long reasoning traces sequentially, creating severe latency for difficult tasks (up to days and weeks). Parallel reasoning offers a natural remedy. However, prior systems primarily focus on Subtask Parallelism, where the model learns to decompose a high-level task into smaller chunks that can be solved independently. This approach overlooks another pervasive form of parallelism: Trial Parallelism, where multiple speculative attempts explore, verify, and aggregate competing hypotheses in parallel. In this paper, we introduce Parason, which reveals and learns both forms of parallelism in LLM reasoning. Our analysis identifies Trial Parallelism as the majority of parallelizable reasoning computation (65.5% in DeepSeek-V4's reasoning steps in HLE), and it becomes increasingly dominant on hard problems. Guided by this taxonomy, Parason converts sequential reasoning traces into structured parallel trajectories with a context-free grammar, then trains models with Parallelism-Aware Group Relative Policy Optimization (PA-GRPO), whose reward jointly balances accuracy, latency, and the two parallelism ratios. At inference time, Parason executes the learned parallel structure through tool calls, translating theoretical savings to real-world wall-clock acceleration. Experiments on mathematical reasoning benchmarks including AIME24 and AIME25 show that Parason achieves an average acceleration about 1.7$\times$ while maintaining competitive accuracy.

cs.AI↗

arXiv2026

Humanize MLE Agents: Multi-Agent Session Alternation Beyond Single-Agent Limits

Existing LLM-based agent approaches for long-horizon machine learning engineering (MLE) often use context engineering, implemented through custom harnesses that manage tools, skills, and accumulated experience. However, we find that as single-session execution lengthens, the agent's reasoning intensity declines alongside slowing progress, summarized as long-horizon deceleration, and the benefits of these custom harnesses vanish as the underlying model improves, summarized as harness staleness. These findings suggest that their benefits may not persist across versions, which motivates a workflow-level approach in which agents build on one another's work across sessions. We model these alternating agent sessions as a context-reset semi-Markov decision process (CR-SMDP) and establish sufficient conditions for alternation to exceed single-agent reference limits, derive an explicit experiment-budget bound, and characterize when session caps preserve this advantage. Guided by this analysis, we introduce Humanize MLE Agents (HMA), which alternates two agents on a shared workspace, limits each session to at most k experiments and resets the context between sessions, without additional task-specific harness or prior knowledge supplements. Across 75 Kaggle tasks in MLE-bench, HMA reaches a state-of-the-art any-medal rate of 78.2%, exceeding the strongest open-source baseline by 8.0 points with a quarter of its time budget. Moreover, among six configurations spanning five heterogeneous agent pairs, HMA improves any-medal rate over the constituent baselines by 6.4 points on average, demonstrating cross-pair generality. A supplementary case on kernel optimization reaches 878 cycles, within 1.6% of the community leaderboard leader, demonstrating applicability beyond MLE.

cs.AI↗

arXiv2026

Humanize Olympiad Agents: A Multi-Session Workflow for Olympiad Problem Solving with Solve, Review, Wiki, Verify

Large language model (LLM) agents are rapidly improving on long-horizon scientific and olympiad tasks. Recent systems have produced a Lean-formalized resolution of the Navier--Stokes problem and autonomous solutions to four open questions in the Erdős Conjectures database. Yet neither system releases both its underlying model and complete inference stack. More broadly, a single agent session remains brittle on difficult problems: early errors can propagate, missing domain knowledge can stall progress, and apparent success can mask a misinterpretation or weakening of the original task. We introduce HOA (Humanize Olympiads Agent), an open, flow-level framework that coordinates multiple agent sessions through three complementary feedback channels. First, an independent reviewer evaluates each candidate and provides actionable feedback. Second, a structured wiki supplies task-specific knowledge that may be missing or unreliable in model pretraining. Third, a verifier checks generated Lean proofs against trusted challenge statements. With GPT-5.6 Sol, HOA achieves perfect scores on the IMO, IPhO, and IOI 2026, near-full marks on the IChO 2026 problem set, and solves all 672 problems (100%) on PutnamBench. On the Lean-Eval leaderboard, HOA ranks first, solving 251 of 304 problems (82.6%). We further evaluate HOA with open models. Using Kimi-K3, HOA achieves full scores on the IMO, IPhO, and IOI, and a gold-level score of 97.1% on the IChO. These results closely match the closed-source GPT-5.6 Sol + /Goal baseline on the IMO and IPhO, while exceeding it on the IChO and IOI, demonstrating that systematic feedback can help bridge substantial differences in base-model capability.

cs.AI↗

arXiv2026

Humanize: Judgement Engineering for Agentic Coding

Agentic coding makes code generation cheap, but reliable completion remains difficult: the agent that writes the code is a weak judge of whether it is done. We present Humanize, a multi-agent orchestration workflow for agentic coding built around judgement engineering: explicit, mechanically enforced decisions at the boundaries between planning, implementation, review, and learning. A human approves a plan contract, a builder agent implements it in rounds, and a reviewer agent from another vendor decides completion; deterministic hooks, not a model, route work between these roles and enforce 72 mechanical gates. Viewed as a Markov chain over repository states, alternating builder and reviewer samples jointly from two models, so a defect survives only if both miss it. We study Humanize through its deployment, 118 public postmortems of real loops, and its applications. Over 68 versions in 108 days, it gathered 1,468 GitHub stars. Applications include a 567-file gem5 build-system migration under upstream review; Kernel Design Agents, which extend the loop with a kernel knowledge base and profiling feedback and placed in the top three of all three Full-Agent tracks of the MLSys 2026 FlashInfer contest; and, through Humanize Olympiad Agents (HOA), full scores in IOI 2026, IMO 2026, IPhO 2026, and IBO 2024, 418.5/437 in IChO 2026 (gold-medal). Humanize also achieves 672/672 on PutnamBench and ranks first (251/303) on Lean-Eval's leaderboard even competiting with professional mathematicians. The postmortems show that independent review catches unsupported builder claims, but stopping remains a key weakness. In reports that separate rounds by phase, two thirds of rounds occurred after implementation was accepted. This evidence is observational, not a controlled comparison of workflows.

cs.AI↗

arXiv2026

Long-WAM: Scaling the Context of World-Action Models

Real-time robot control demands enough visual history to infer motion and task progress, but processing that history can delay action. We present Long-WAM, a model-system framework for scaling the context of causal world-action models under real-time control constraints. Our central finding is that access to history is not the same as using it: longer histories pay off far more when the video foundation is pretrained autoregressively (AR). We first learn causal prediction from robot and egocentric videos without action labels, then preserve this history-to-future structure during world-action adaptation. On RoboCasa GR-1, increasing context from 0.0 to 19.2 seconds raises success from 63.3% to 78.7%, whereas a bidirectionally pretrained initialization shows no net gain; robot-domain AR pretraining further raises peak success on GR-1 and LIBERO-Long. Long-WAM also achieves the best results among compared methods on LIBERO-Long, RoboTwin 2.0, and DOMINO. Streaming observation encoding, asynchronous execution, and hardware-specific acceleration enable deployment on RTX 5090, DGX Spark, and Jetson AGX Thor without dropping future prediction; on RTX 5090, each action chunk, including future-video latent prediction, takes 107.4 ms. Real-time deployment on Unitree G1 and YAM supports dynamic and long-horizon manipulation, including 95% success on dynamic cup stacking, where Pi0.5 and Fast-WAM succeed in none of 20 trials. As a memory-informed executor, Long-WAM also complements higher-level planning in composite tasks.

cs.RO↗

arXiv2026

Physis-Lang: Self-Evolving Language as a Physical Representation for Video World Model

Video world models are expected to predict how the physical world evolves, yet they often produce visually plausible videos that violate basic physical principles. Existing approaches commonly assume that natural language is insufficient to represent the physical knowledge required for reliable generation, and therefore introduce additional visual, latent, numerical, or planning-based signals. We revisit this assumption and introduce Physis-Lang, a self-evolving framework that treats physical language as a shared and optimizable representation across data curation, model training, and video generation. Physis-Lang represents physical processes through language that describes their relevant entities, causes, interactions, governing principles, temporal evolution, and effects. To improve this representation, we construct PhysCapBench, which decomposes physical processes into atomic assertions and evaluates captions using recall and precision. An agentic loop iteratively analyzes assertion-level errors and refines the instruction used to produce physical captions. Physis-Lang further converts model deficiencies into textual descriptions and uses language-guided retrieval to identify visually diverse videos that cover missing physical processes. Experiments on four widely used physical video benchmarks with Wan and Cosmos backbones demonstrate consistent improvements in physical plausibility. Notably, starting from open-source Cosmos3-Nano backbones, our Physis-Lang-enhanced models surpass the leading proprietary Veo 3.1 model.

cs.CV↗

arXiv2026

Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE

Modern LLMs are increasingly deployed in long-context applications such as retrieval-augmented generation, repository-level coding, and agentic workflows whose accumulated reasoning and tool traces routinely push the input an order of magnitude past the pretraining window, making zero-shot context extension the dominant deployment path for open-weight checkpoints. The dominant zero-shot methods (YaRN, Self-Extend, DCA) fix a single rescaling factor up front, so an aggressive factor sacrifices short-context fidelity while a conservative one breaks down at long contexts; recent length-aware variants adapt the mapping, but with a fitted or distance-dependent schedule. We propose Jet-Long, a tuning-free zero-shot method that pairs a local RoPE-faithful window with a long-range window whose rescaling factor adapts dynamically to the current sequence length via a parameter-free analytic schedule, recovering the base model exactly at short inputs while extrapolating cleanly at long ones. An inclusion-exclusion attention merge and on-the-fly RoPE correction enable a fused CuTe implementation. On H100 at 64K-128K, prefill retains 83-88% of FlashAttention-3 throughput across the evaluated Qwen3 sizes and 88-93% of a matched CuTe control; Qwen3-8B single-batch generation reaches 1.04-1.08 times FlashAttention-3 throughput. On Qwen3-1.7B/4B/8B up to 128K context, Jet-Long leads RULER by +4.79/+2.18/+2.03 percentage points over the strongest baseline at 1.7B/4B/8B, achieves the best overall accuracy on HELMET-RAG (a benchmark identified by HELMET as the most efficient predictor of downstream long-context performance) and attains the lowest PG-19 perplexity. Additional evaluations cover Meta-Llama-3-8B, post-trained Qwen3 checkpoints, and the hybrid Jet-Nemotron architecture, supporting broader applicability without retraining. The local-window hyperparameter remains robust across the tested settings.

cs.LG↗

arXiv2026

SoL-Refiner: Speed-of-Light One-Step Refinement for High-Resolution Video

High-resolution video generation is expensive, as its cost grows rapidly with the number of spatiotemporal tokens. A practical alternative first generates a lower-resolution video and then applies a refiner, but conventional multi-step refinement introduces a second sampling bottleneck. We present SoL-Refiner, a one-step video refiner that transforms low-resolution model outputs into 4K videos with a single denoising step. Our three-stage recipe combines high-resolution continual training, reinforcement learning (RL) post-training, and a final one-step distillation. We introduce Refiner-Bench, a video refinement benchmark constructed from the outputs of different video generators, and use a shared-input protocol to compare refiners at approximately 2K output resolution. At 2K, the one-step SoL-Refiner outperforms all external refiners on the VBench and UniPercept averages, while at $3840\!\times\!2176$ it improves both metrics over the three-step LTX-2.3 Refiner. With the complete acceleration stack, SoL-Refiner achieves an $8.91\times$ speedup in refinement latency over the same baseline in our 2K latency setting.

cs.CV↗

arXiv2026

LongLive-Plug: Once-for-All Distillation for Video Generation

Video diffusion models are increasingly developed into specialized models for diverse downstream tasks, and this development often includes a distillation stage, for example to accelerate sampling or to improve long-video generation. This stage is typically repeated for every specialized model. We introduce LongLive-Plug, a once-for-all distillation framework that learns reusable capabilities as LoRAs on a base model for training-free, plug-and-play deployment to compatible downstream models. These capabilities include single-pass classifier-free guidance, few-step sampling, and long-context error correction for autoregressive generation. The adapters remain reusable even when downstream models add conditioning branches, expand output channels. Despite training at a fixed guidance scale, our dedicated CFG LoRA provides text guidance control through its inference weight. Combining it with a few-step LoRA simultaneously preserves few-step generation and CFG controllability on downstream tasks. We verify training-free deployment on 54 downstream models across three backbone families and eight task categories, including world modeling, robotics, editing, and multimodal generation. The approach may support additional compatible models. Each capability can thus be distilled once per backbone family and reused without per-target retraining.

cs.CV↗

arXiv2026

LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization

Recent LLMs increasingly adopt hybrid designs that replace standard attention with linear attention, such as Gated DeltaNet (GDN) and Kimi Delta Attention (KDA). Although they compress the context into a fixed-size recurrent state and substantially reduce the cost of long-context processing, repeatedly reading and updating that state remains a major inference bottleneck. Quantization offers a natural way to reduce this cost, but can significantly degrade model quality, due to the accumulation of rounding errors and the presence of outlier rows and columns in the state. To address these challenges, we propose LeapQuant, a training-free method that achieves near-lossless performance under 8-bit recurrent-state quantization. First, to mitigate error accumulation, we propose per-window quantization, which leaps over a window of tokens and quantizes the state only once at its end. Within a window, outputs are computed from the fixed low-bit state together with high-precision buffered updates. Second, to reduce the error introduced by each quantization, LeapQuant retains the state's largest outliers as a few high-precision Compensator Tokens, which share the update path of real tokens. We then smooth the remaining residual before quantization to further reduce the error. Comprehensive experiments across the Qwen, Kimi, and GLM model families show that LeapQuant substantially reduces memory and compute costs during inference. With accuracy comparable to the FP32 baseline, it achieves average speedups of 2.05--3.70$\times$ at the kernel level and 1.47$\times$ for end-to-end inference on NVIDIA B200, RTX PRO 6000, and RTX 5090 GPUs.

cs.LG↗

arXiv2026

Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge

Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.

cs.AI↗

arXiv2026

d-OPD: Future-Aware On-Policy Distillation for Block Diffusion Language Models

Large language models (LLMs) typically generate text autoregressively (AR), predicting one token at a time. Block diffusion language models (dLLMs) instead generate blocks sequentially while denoising multiple tokens in parallel within each block, offering a promising way to accelerate generation. Rather than training such models from scratch, recent work adapts strong pretrained AR models into block dLLMs through distillation. On-policy distillation (OPD) has been widely used for LLM training because it supervises the student on states generated by its current policy, rather than only on fixed offline trajectories. By training on the states the student actually visits, it reduces the mismatch between training and generation and can provide more relevant supervision as the student evolves. Recent work has extended this idea to AR-to-block-diffusion conversion. However, this setting introduces a fundamental mismatch in supervision: the block-diffusion student and the causal AR teacher condition on different information at the same training state. The student predicts from the entire partially denoised block, including visible future context, whereas the standard AR teacher target is defined only from the causal prefix. As a result, the teacher distribution used for distillation is not fully aligned with the information available to the student. We therefore introduce d-OPD, a future-aware on-policy distillation method that corrects the AR teacher distribution to better align with the student-visible state by incorporating visible future information within each block, providing supervision that better matches the information used by the student. Across Qwen3 models from 0.6B to 8B, d-OPD improves the six-benchmark average by up to $4.0$ points over OPDLM and reduces training time by $1.35$-$1.58\times$. The code is available at https://github.com/mit-han-lab/d-OPD.

cs.LG↗

arXiv2026

Lightning OPD: Efficient Post-Training for Large Reasoning Models with Offline On-Policy Distillation

On-policy distillation (OPD) is an effective post-training paradigm for large language models but requires a live teacher server throughout training, resulting in substantial infrastructure overhead. We investigate whether OPD can be performed offline by precomputing teacher log-probabilities once over SFT rollouts and reusing them during training. We find that naively doing so fails to reliably match standard OPD, and trace the root cause to a previously overlooked condition we term teacher consistency, requiring that the same teacher be used for both supervised fine-tuning and OPD. Violating this condition introduces a gradient bias that degrades performance for both offline and online OPD. Building on this insight, we propose Lightning OPD, an offline on-policy distillation framework that enforces teacher consistency and eliminates the need for a live teacher server entirely. We prove that, under teacher consistency, Lightning OPD shares the same optimum as standard OPD, with bounded gradient discrepancy and an implicit regularization effect that helps prevent policy drift. Experiments on math reasoning and code generation show that Lightning OPD achieves comparable performance to standard OPD while delivering 4.0x higher training efficiency. Starting from an SFT-initialized Qwen3-8B-Base model, Lightning OPD reaches 69.9% on AIME 2024 in just 30 GPU hours. Lightning OPD further scales to MoE architectures, training Qwen3-30B-A3B to 71.0% on AIME 2024 on a single 8xH100 node, substantially lowering the barrier for academic research on LLM post-training. Our code is released at https://github.com/jet-ai-projects/Lightning-OPD.

cs.LG↗

arXiv2026

SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness

As coding agents move from supervised code completion to unattended, around-the-clock exploration, their work expands from isolated predictions into long trajectories of reasoning, tool use, and feedback. Token efficiency therefore becomes important for scaling recursive self-improvement. We take an RSI-inspired approach at the harness layer, scaling auto-research loops across increasingly numerous and diverse environments for harness rollouts. At this scale, the process yields reusable improvements that transfer beyond their development setting, moving automated harness discovery toward production-level outcomes. Four mechanisms survive selection and form SoL-Pi, spanning action execution, context compaction, observation handling, and delegated reading. On the 51-task EdgeBench evaluation, SoL-Pi achieves performance comparable to Pi across GPT-5.6 Sol and Opus 5 while reducing recorded token traffic by 44.7-49.0% and API cost by about one third. In other words, estimated hourly savings are \$8.75-\$13.50 relative to native Codex and Claude Code harnesses, and \$4.36-\$5.71 relative to Pi.

cs.AI↗