DiscoLoop adds a decoded token-embedding channel to looped transformers, fixing a representation mismatch that limited implicit multi-hop reasoning and improving OOD generalization.
hub Canonical reference
Qwen3-Coder-Next Technical Report
Canonical reference. 88% of citing Pith papers cite this work as background.
40
Pith papers citing it
Background
88% of classified citations
abstract
We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. Qwen3-Coder-Next is an 80-billion-parameter model that activates only 3 billion parameters during inference, enabling strong coding capability with efficient inference. In this work, we explore how far strong training recipes can push the capability limits of models with small parameter footprints. To achieve this, we perform agentic training through large-scale synthesis of verifiable coding tasks paired with executable environments, allowing learning directly from environment feedback via mid-training and reinforcement learning. Across agent-centric benchmarks including SWE-Bench and Terminal-Bench, Qwen3-Coder-Next achieves competitive performance relative to its active parameter count. We release both base and instruction-tuned open-weight versions to support research and real-world coding agent development.
hub tools
citation-role summary
background 7
baseline 1
citation-polarity summary
years
2026 40representative citing papers
CLI-Universe synthesizes a verified 6K dataset of terminal-agent tasks that, when used to fine-tune Qwen3-32B, reaches 33.4% on Terminal-Bench 2.0 and sets a new open-source SOTA for models at or below 32B parameters.
Incomplete constrainers in constrained decoding push LLMs into low-probability program regions, making unconstrained decoding outperform constrained decoding on functional correctness across seven models and three benchmarks.
StaminaBench evaluates coding agents over 100 procedurally generated change requests to a REST API, finding that tested models fail within 5-6 turns without feedback but improve up to 12x with test feedback and good harnesses.
WhiFlash introduces token-level cross-paradigm routing between autoregressive and diffusion drafting models, with cache optimizations, to raise acceptance lengths and deliver up to 69.6% throughput gains over EAGLE-3.
AssertLLM2 introduces a benchmark of 83 designs supporting bug-prevention and bug-hunting assertion generation tasks with evaluation across syntactic, formal, coverage, and mutation-based metrics.
SpecBench shows frontier coding agents saturate visible test suites but exhibit persistent reward hacking on held-out tests, with the gap growing 28 percentage points per tenfold increase in code size.
CopT reverses CoT by eliciting a draft answer first then using continuous-embedding contrastive verification and on-policy thinking to reflect and correct, yielding up to 23% higher accuracy and 57% fewer tokens without training.
AssayBench is a new gene-ranking benchmark for phenotypic CRISPR screens that shows zero-shot generalist LLMs outperform both biology-specific LLMs and trainable baselines on adjusted nDCG.
Code LLMs generate substantially worse comments outside English, and no tested automatic metric or LLM judge reliably matches human assessment of those outputs.
DEONTICBENCH is a new benchmark of 6,232 deontic reasoning tasks from U.S. legal domains where frontier LLMs reach only ~45% accuracy and symbolic Prolog assistance plus RL training still fail to solve tasks reliably.
On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.
Minos uses a two-tiered multi-agent architecture with retrieval-augmented reasoning and FSM-coordinated agents to reconstruct attack scenarios from provenance data, reporting 0.92 recall and 0.64 precision on 14 scenarios.
G2PO transforms linear trajectories into graphs, aggregates identical states for lower-variance value estimates, and uses edge-centric TD standardization, reporting up to 22.2% gains over GRPO on WebShop, ALFWorld, and AppWorld.
Code evolution produces link prediction algorithms with average AUC of 0.915 versus 0.783 for human-designed methods across 580 networks, with better scalability.
The authors developed an evaluation framework that generates 1000 tasks from 500 real-world agent skills, applies instruction-following and goal-completion rubrics, and benchmarks 19 proprietary and open-source model configurations.
PEEK maintains a constant-sized context map via a programmable cache policy to give LLM agents persistent orientation knowledge about recurring external contexts, yielding 6-34% gains and lower cost than prior prompt-learning methods.
OScaR mitigates token norm imbalance via canalized rotation and omni-token scaling to enable near-lossless INT2 KV cache quantization with up to 3x decoding speedup and 5.3x memory reduction.
CRANE applies magnitude thresholding, a Conservative Taylor Gate, and Graduated Sigmoidal Projection to the Thinking-Instruct delta to improve code agent pass rates on Roo-Eval, SWE-bench-Verified, and Terminal-Bench while preserving efficiency.
DAgger-style training with turn-level policy interpolation raises 4B and 8B LLM agents to 27.3% and 29.8% on SWE-bench Verified, beating several larger published systems.
Priming transfers knowledge from pre-trained Transformers to hybrid SSM-attention models, recovering performance with minimal additional tokens and showing Gated KalmaNet outperforming Mamba-2 on long-context reasoning at 32B scale.
SynConfRoute routes code completions using syntax validation and token confidence, improving pass@1 by up to 31% on hard tasks and reducing accelerator usage by 58% versus always using the largest model.
Terminus-4B, a Qwen3-4B post-trained with SFT and rubric-based GRPO, matches or beats frontier models as a terminal-execution subagent while cutting main-agent tokens ~30% on SWE-Bench-style tasks.
EVICT adaptively truncates draft trees in MoE speculative decoding by combining drafter signals with profiled costs to retain only cost-effective prefixes, delivering up to 2.35x speedup over autoregressive decoding.
citing papers explorer
-
DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning
DiscoLoop adds a decoded token-embedding channel to looped transformers, fixing a representation mismatch that limited implicit multi-hop reasoning and improving OOD generalization.
-
CLI-Universe: Towards Verifiable Task Synthesis Engine for Terminal Agents
CLI-Universe synthesizes a verified 6K dataset of terminal-agent tasks that, when used to fine-tune Qwen3-32B, reaches 33.4% on Terminal-Bench 2.0 and sets a new open-source SOTA for models at or below 32B parameters.
-
The Alignment Problem in Constrained Code Generation
Incomplete constrainers in constrained decoding push LLMs into low-probability program regions, making unconstrained decoding outperform constrained decoding on functional correctness across seven models and three benchmarks.
-
StaminaBench: Stress-Testing Coding Agents over 100 Interaction Turns
StaminaBench evaluates coding agents over 100 procedurally generated change requests to a REST API, finding that tested models fail within 5-6 turns without feedback but improve up to 12x with test feedback and good harnesses.
-
WhiFlash: Accelerating Speculative Decoding with Token-Level Cross-Paradigm Routing
WhiFlash introduces token-level cross-paradigm routing between autoregressive and diffusion drafting models, with cache optimizations, to raise acceptance lengths and deliver up to 69.6% throughput gains over EAGLE-3.
-
AssertLLM2: A Comprehensive LLM Benchmark for Assertion Generation from Design Specifications
AssertLLM2 introduces a benchmark of 83 designs supporting bug-prevention and bug-hunting assertion generation tasks with evaluation across syntactic, formal, coverage, and mutation-based metrics.
-
SpecBench: Measuring Reward Hacking in Long-Horizon Coding Agents
SpecBench shows frontier coding agents saturate visible test suites but exhibit persistent reward hacking on held-out tests, with the gap growing 28 percentage points per tenfold increase in code size.
-
CopT: Contrastive On-Policy Thinking with Continuous Spaces for General and Agentic Reasoning
CopT reverses CoT by eliciting a draft answer first then using continuous-embedding contrastive verification and on-policy thinking to reflect and correct, yielding up to 23% higher accuracy and 57% fewer tokens without training.
-
AssayBench: An Assay-Level Virtual Cell Benchmark for LLMs and Agents
AssayBench is a new gene-ranking benchmark for phenotypic CRISPR screens that shows zero-shot generalist LLMs outperform both biology-specific LLMs and trainable baselines on adjusted nDCG.
-
Evaluating Non-English Developer Support in Machine Learning for Software Engineering
Code LLMs generate substantially worse comments outside English, and no tested automatic metric or LLM judge reliably matches human assessment of those outputs.
-
DeonticBench: A Benchmark for Reasoning over Rules
DEONTICBENCH is a new benchmark of 6,232 deontic reasoning tasks from U.S. legal domains where frontier LLMs reach only ~45% accuracy and symbolic Prolog assistance plus RL training still fail to solve tasks reliably.
-
From Execution to Education: A Bloom-Aligned Framework for Measuring Educational Control in LLMs
On 2,520 programming tasks, matched Qwen general and coder models reliably raise Bloom cognitive demand but fail to lower it, so execution skill does not imply educational control.
-
Minos: A Multi-Agent Collaborative Framework for Provenance-Based Backward Tracking
Minos uses a two-tiered multi-agent architecture with retrieval-augmented reasoning and FSM-coordinated agents to reconstruct attack scenarios from provenance data, reporting 0.92 recall and 0.64 precision on 14 scenarios.
-
Group-Graph Policy Optimization for Long-Horizon Agentic Reinforcement Learning
G2PO transforms linear trajectories into graphs, aggregates identical states for lower-variance value estimates, and uses edge-centric TD standardization, reporting up to 22.2% gains over GRPO on WebShop, ALFWorld, and AppWorld.
-
Code evolution for link prediction in complex networks
Code evolution produces link prediction algorithms with average AUC of 0.915 versus 0.783 for human-designed methods across 580 networks, with better scalability.
-
A Framework for Evaluating Agentic Skills at Scale
The authors developed an evaluation framework that generates 1000 tasks from 500 real-world agent skills, applies instruction-following and goal-completion rubrics, and benchmarks 19 proprietary and open-source model configurations.
-
PEEK: Context Map as an Orientation Cache for Long-Context LLM Agents
PEEK maintains a constant-sized context map via a programmable cache policy to give LLM agents persistent orientation knowledge about recurring external contexts, yielding 6-34% gains and lower cost than prior prompt-learning methods.
-
OScaR: The Occam's Razor for Extreme KV Cache Quantization in LLMs and Beyond
OScaR mitigates token norm imbalance via canalized rotation and omni-token scaling to enable near-lossless INT2 KV cache quantization with up to 3x decoding speedup and 5.3x memory reduction.
-
CRANE: Constrained Reasoning Injection for Code Agents via Nullspace Editing
CRANE applies magnitude thresholding, a Conservative Taylor Gate, and Graduated Sigmoidal Projection to the Thinking-Instruct delta to improve code agent pass rates on Roo-Eval, SWE-bench-Verified, and Terminal-Bench while preserving efficiency.
-
Revisiting DAgger in the Era of LLM-Agents
DAgger-style training with turn-level policy interpolation raises 4B and 8B LLM agents to 27.3% and 29.8% on SWE-bench Verified, beating several larger published systems.
-
Priming: Hybrid State Space Models From Pre-trained Transformers
Priming transfers knowledge from pre-trained Transformers to hybrid SSM-attention models, recovering performance with minimal additional tokens and showing Gated KalmaNet outperforming Mamba-2 on long-context reasoning at 32B scale.
-
SynConfRoute: Syntax-Aware Routing for Efficient Code Completion with Small CodeLLMs
SynConfRoute routes code completions using syntax validation and token confidence, improving pass@1 by up to 31% on hard tasks and reducing accelerator usage by 58% versus always using the largest model.
-
Terminus-4B: Can a Smaller Model Replace Frontier LLMs at Agentic Execution Tasks?
Terminus-4B, a Qwen3-4B post-trained with SFT and rubric-based GRPO, matches or beats frontier models as a terminal-execution subagent while cutting main-agent tokens ~30% on SWE-Bench-style tasks.
-
Making Every Verified Token Count: Adaptive Verification for MoE Speculative Decoding
EVICT adaptively truncates draft trees in MoE speculative decoding by combining drafter signals with profiled costs to retain only cost-effective prefixes, delivering up to 2.35x speedup over autoregressive decoding.
-
From Natural Language to Verified Code: Toward AI Assisted Problem-to-Code Generation with Dafny-Based Formal Verification
Open-weight LLMs reach 81-91% success generating formally verified Dafny code for complex algorithmic problems when given structural signatures and self-healing verifier feedback.
-
VERITAS: A Multi-Agent Co-Scientist for Verifiable Image-Derived Hypothesis Testing
A four-phase multi-agent co-scientist tests natural-language hypotheses on cardiac and glioma MRI and labels outcomes Supported, Refuted, Underpowered, or Invalid with an executable evidence trail.
-
REAgent: Requirement-Driven LLM Agents for Software Issue Resolution
REAgent improves LLM patch generation for software issues by 17.4% on average through automated construction, quality checking, and iterative refinement of structured issue-oriented requirements.
-
From SWE-ZERO to SWE-HERO: Execution-free to Execution-based Fine-tuning for Software Engineering Agents
A two-stage SFT pipeline distills execution-free then execution-based trajectories from a 480B model into smaller Qwen2.5-Coder agents, yielding 62.2% resolution on SWE-bench Verified and 44.1% zero-shot on the multilingual version.
-
BlockPilot: Instance-Adaptive Policy Learning for Diffusion-based Speculative Decoding
BlockPilot is an instance-adaptive policy that predicts optimal block size from the prefilling representation for diffusion speculative decoding, reporting 5.92 acceptance length and 4.20x speedup on Qwen3-4B.
-
Piper: A Programmable Distributed Training System
Piper decouples user-defined distributed training strategies from runtime execution using transformations on a unified global training DAG IR, achieving parity on ZeRO and gains on composed strategies like DualPipe.
-
Rethinking Continual Experience Internalization for Self-Evolving LLM Agents
Existing methods for turning LLM interaction experience into parametric skills collapse over multiple iterations; principle-level experience, step-wise injection, and off-policy teacher distillation yield more stable continual learning.
-
Code as a Weapon: A Consensus-Labeled Prompt Bank for Measuring Coding-Model Compliance with Malicious-Code Requests
Consolidates eight corpora into a 6,671-prompt bank with five-judge consensus labels separating executable malicious code requests (4,748) from harmful security knowledge requests (1,923), achieving Fleiss' kappa 0.767.
-
$\boldsymbol{f}$-OPD: Stabilizing Long-Horizon On-Policy Distillation with Freshness-Aware Control
f-OPD decomposes on-policy distillation drift into rollout and supervision components, then applies a sample-level freshness score to adaptively limit stale data influence and stabilize long-horizon agent training.
-
A Validated Prompt Bank for Malicious Code Generation: Separating Executable Weapons from Security Knowledge in 1,554 Consensus-Labeled Prompts
The paper releases a 1,554-prompt consensus-labeled bank separating executable malicious code requests from security knowledge requests, validated by five-model majority labeling with Fleiss' kappa of 0.876.
-
LayerBoost: Layer-Aware Attention Reduction for Efficient LLMs
LayerBoost selectively replaces or removes attention in non-critical transformer layers to cut inference latency up to 68% while recovering quality via brief distillation.
-
Bridging the Gap between User Intent and LLM: A Requirement Alignment Approach for Code Generation
REA-Coder improves LLM code generation by iteratively aligning requirements with model understanding and verifying outputs against the aligned spec.
-
DAR: Deontic Reasoning with Agentic Harnesses
DAR lets LLMs interact dynamically with statutes for deontic reasoning, improving results on hard DeonticBench subsets but with uneven gains and higher token use for weaker models.
-
CLI-Anything: Towards Agent-Native Computer Use
CLI-Anything advocates transforming applications into CLI-based protocols for agent-native interaction and introduces the CLI-Hub platform to support this shift away from GUI agents.
-
PASK: Toward Intent-Aware Proactive Agents with Long-Term Memory
PASK introduces the DD-MM-PAS paradigm for streaming proactive agents with intent-aware detection, hybrid memory modeling, and a new real-world benchmark where the IntentFlow model matches top LLMs on latency while finding deeper intents.
- PDE-Agents: An LLM-Orchestrated Multi-Agent Framework for Automated Finite Element Simulations with Knowledge Graph-Augmented Reasoning