Selected work

Representative Papers

SWE-World: Building Software Engineering Agents in Docker-Free Environments

Feb 03, 2026

This work proposes SWE-World, a novel framework that eliminates the need for physical Docker containers in software engineering agent training. Traditional approaches rely on containerized execution to obtain feedback, incurring substantial resource overhead and limiting scalability. In contrast, SWE-World leverages large language models to construct a simulated environment that predicts intermediate execution states and test outcomes based on real-world interaction data. This enables supervised fine-tuning (SFT), reinforcement learning (RL), and test-time multi-trajectory selection (TTS) entirely within a virtual setting. Evaluated on SWE-bench Verified, the framework boosts the pass rate of Qwen2.5-Coder-32B from 6.2% to 68.2%, demonstrating its effectiveness and scalability.

1 citationsRead paper

Evo-Bench: Can Language Models Improve Agent Harness?

Aug 09, 2026

Existing evaluation methods struggle to independently assess an agent’s ability to autonomously optimize its operational harness, often conflated by the underlying model’s performance and lacking characterization of long-term evolution. To address this, this work proposes Evo-Bench—the first benchmark specifically designed to evaluate language models’ capacity for self-evolving their harnesses—spanning search, office, and general domains. By leveraging a harness-guided construction framework, auxiliary task evolution, and sensitivity-aware hierarchical partitioning, Evo-Bench enables systematic and decoupled assessment of harness evolution capabilities. Experiments show that leading models achieve up to a 16.6-point improvement on Evo-Bench, approaching human-designed optimal baselines, with strong performance in general and search tasks, though gaps remain in office tasks requiring specific procedural knowledge. These results validate both the high transferability of synthesized harnesses and the effectiveness of the proposed evaluation framework.

0 citationsRead paper

Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Mode

Jul 24, 2026

This work addresses the challenge of enabling small-scale models to achieve strong agentic capabilities under strict parameter constraints by proposing a compact, general-purpose agent model with only 3 billion non-embedding parameters. The model is trained from scratch on 28 trillion tokens of diverse real and synthetic agent trajectories and employs an innovative Looped Transformer architecture that reuses layer stacks to increase effective capacity without expanding parameter count. Enhanced through hybrid-mode RLHF, length-controlled reinforcement learning, and a dual reward mechanism balancing process and outcome, the model demonstrates significantly improved performance in multi-task reasoning, code generation, and tool use. It outperforms larger counterparts such as Qwen3.5-9B and Gemma4-12B across multiple agent benchmarks, excelling in mathematical, programming, scientific reasoning, and alignment tasks, thereby serving as an efficient and lightweight local personal assistant.

0 citationsRead paper

HierarchicalKV: A GPU Hash Table with Cache Semantics for Continuous Online Embedding Storage

Mar 17, 2026

This work addresses severe memory inefficiency in conventional GPU hash tables when embedding tables exceed the capacity of a single GPU’s high-bandwidth memory (HBM), as these structures retain all key-value pairs regardless of access patterns. To overcome this limitation, the authors propose HierarchicalKV—the first GPU hash table that treats caching semantics as a first-class operation. It replaces traditional dictionary semantics with a policy-driven eviction mechanism that either updates entries in place or rejects insertions, thereby avoiding costly rehashing and overflow failures. Key innovations include cache-line-aligned buckets, inline score-driven upserts, dynamic dual-bucket selection, three-level concurrency control, and a hierarchical key-value separation architecture. Evaluated on an NVIDIA H100 NVL, HierarchicalKV achieves up to 3.9 billion key-value operations per second, maintains load factors between 0.50 and 1.00 with less than 5% throughput variation, outperforms WarpCore by 1.4×, and surpasses indirect-addressing baselines by 2.6–9.4×, with integration already adopted in multiple open-source recommendation frameworks.

0 citationsRead paper

Measuring and Mitigating Post-hoc Rationalization in Reverse Chain-of-Thought Generation

Feb 16, 2026

This work addresses the post-hoc rationalization problem in existing Reverse Chain-of-Thought (Reverse CoT) methods, where exposure to the ground-truth answer biases reasoning by serving as a cognitive anchor. The study presents the first systematic quantification of this anchoring effect through a tripartite evaluation framework encompassing lexical, entropy-based, and probabilistic metrics. Drawing on the ironic process theory from cognitive psychology, the authors propose a novel Structured Skeleton-guided Reasoning (SSR) paradigm that first generates an answer-agnostic functional skeleton and subsequently constructs a complete reasoning chain based on this scaffold. They further introduce a distillation-based fine-tuning strategy, SSR-D. Experimental results demonstrate that SSR-D achieves up to a 10% performance gain over semantic suppression baselines across multiple open-ended reasoning benchmarks and exhibits strong out-of-distribution generalization capabilities.

0 citationsRead paper

Recent publications

Latest Papers

Evo-Bench: Can Language Models Improve Agent Harness?

Aug 09, 2026

Existing evaluation methods struggle to independently assess an agent’s ability to autonomously optimize its operational harness, often conflated by the underlying model’s performance and lacking characterization of long-term evolution. To address this, this work proposes Evo-Bench—the first benchmark specifically designed to evaluate language models’ capacity for self-evolving their harnesses—spanning search, office, and general domains. By leveraging a harness-guided construction framework, auxiliary task evolution, and sensitivity-aware hierarchical partitioning, Evo-Bench enables systematic and decoupled assessment of harness evolution capabilities. Experiments show that leading models achieve up to a 16.6-point improvement on Evo-Bench, approaching human-designed optimal baselines, with strong performance in general and search tasks, though gaps remain in office tasks requiring specific procedural knowledge. These results validate both the high transferability of synthesized harnesses and the effectiveness of the proposed evaluation framework.

0 citationsRead paper

Nanbeige4.2-3B: Unlocking Agentic Capabilities in a Compact Mode

Jul 24, 2026

This work addresses the challenge of enabling small-scale models to achieve strong agentic capabilities under strict parameter constraints by proposing a compact, general-purpose agent model with only 3 billion non-embedding parameters. The model is trained from scratch on 28 trillion tokens of diverse real and synthetic agent trajectories and employs an innovative Looped Transformer architecture that reuses layer stacks to increase effective capacity without expanding parameter count. Enhanced through hybrid-mode RLHF, length-controlled reinforcement learning, and a dual reward mechanism balancing process and outcome, the model demonstrates significantly improved performance in multi-task reasoning, code generation, and tool use. It outperforms larger counterparts such as Qwen3.5-9B and Gemma4-12B across multiple agent benchmarks, excelling in mathematical, programming, scientific reasoning, and alignment tasks, thereby serving as an efficient and lightweight local personal assistant.

0 citationsRead paper

HierarchicalKV: A GPU Hash Table with Cache Semantics for Continuous Online Embedding Storage

Mar 17, 2026

This work addresses severe memory inefficiency in conventional GPU hash tables when embedding tables exceed the capacity of a single GPU’s high-bandwidth memory (HBM), as these structures retain all key-value pairs regardless of access patterns. To overcome this limitation, the authors propose HierarchicalKV—the first GPU hash table that treats caching semantics as a first-class operation. It replaces traditional dictionary semantics with a policy-driven eviction mechanism that either updates entries in place or rejects insertions, thereby avoiding costly rehashing and overflow failures. Key innovations include cache-line-aligned buckets, inline score-driven upserts, dynamic dual-bucket selection, three-level concurrency control, and a hierarchical key-value separation architecture. Evaluated on an NVIDIA H100 NVL, HierarchicalKV achieves up to 3.9 billion key-value operations per second, maintains load factors between 0.50 and 1.00 with less than 5% throughput variation, outperforms WarpCore by 1.4×, and surpasses indirect-addressing baselines by 2.6–9.4×, with integration already adopted in multiple open-source recommendation frameworks.

0 citationsRead paper

Measuring and Mitigating Post-hoc Rationalization in Reverse Chain-of-Thought Generation

Feb 16, 2026

This work addresses the post-hoc rationalization problem in existing Reverse Chain-of-Thought (Reverse CoT) methods, where exposure to the ground-truth answer biases reasoning by serving as a cognitive anchor. The study presents the first systematic quantification of this anchoring effect through a tripartite evaluation framework encompassing lexical, entropy-based, and probabilistic metrics. Drawing on the ironic process theory from cognitive psychology, the authors propose a novel Structured Skeleton-guided Reasoning (SSR) paradigm that first generates an answer-agnostic functional skeleton and subsequently constructs a complete reasoning chain based on this scaffold. They further introduce a distillation-based fine-tuning strategy, SSR-D. Experimental results demonstrate that SSR-D achieves up to a 10% performance gain over semantic suppression baselines across multiple open-ended reasoning benchmarks and exhibits strong out-of-distribution generalization capabilities.

0 citationsRead paper

Nanbeige4.1-3B: A Small General Model that Reasons, Aligns, and Acts

Feb 13, 2026

This work proposes a 3-billion-parameter universal language model that unifies strong reasoning, human preference alignment, code generation, and complex tool usage within a compact architecture. The model integrates pointwise and pairwise reward modeling to enhance both reasoning and alignment capabilities, employs a complexity-aware reinforcement learning reward mechanism to optimize code generation, and leverages multi-turn synthetic data combined with turn-level supervised training to support extended tool interactions. As the first open-source small-scale system integrating agentic behavior, code synthesis, and general-purpose reasoning, it significantly outperforms comparable models such as Nanbeige4-3B-2511 and Qwen3-4B across multiple benchmarks, and even surpasses the much larger 30B-parameter Qwen3-30B-A3B, demonstrating that small models can achieve both breadth and depth in capability.

0 citationsRead paper