6arXiv cs.LG (Machine Learning)·
MindForge pipeline fine-tunes small models for whole-life-cycle software engineering via source-free program synthesis
MindForge is an automated pipeline that converts open-source command-line programs into source-free training environments exposing only compiled executables and documentation, enabling training data generation for from-scratch program synthesis. Using GLM-5.2 as a teacher agent, the authors fine-tune Qwen3.6-27B on synthesized trajectories, raising its ProgramBench pass rate from 37.98% to 49.51% and achieving gains across seven held-out benchmarks including SWE-bench Verified (+5.04) and RepoZero-C2Rust (+31.00). The work addresses a gap in coding agent training infrastructure by spanning the full software engineering life cycle rather than single-phase tasks. The result is notable for achieving frontier-comparable performance on a 27B model through targeted data curation.
Related guides (4)
Related events (8)
PROVE framework trains LLMs for multi-step tool use via stateful MCP environments and programmatic rewards
Researchers introduce PROVE (Programmatic Rewards On Verified Environments), a framework for training LLMs to orchestrate multi-step tool calls using reinforcement learning. The system includes a library of 20 stateful MCP servers with 343 tools, an automated data synthesis pipeline that grounds training queries in live server state, and a multi-component programmatic reward function requiring no judge model. Training four models (Qwen3-4B, Qwen3-8B, Qwen2.5-7B, Granite-4.1-8B) with ~13K examples yields gains of up to +10.2 on BFCL Multi-Turn, +6.8 on tau2-bench, and +6.5 on T-Eval, demonstrating consistent improvements in multi-step tool orchestration.
Frontier coding agents use metaprogramming to handle esoteric programming languages
A new arXiv paper evaluates six LLM-based coding agents on four esoteric programming languages (including Brainfuck and Befunge-98), finding that the strongest agents—Claude Opus 4.6 and GPT-5.4 xhigh—often avoid writing the target language directly, instead generating it via Python metaprograms. Forbidding this strategy causes large performance drops, and text guidance alone does not transfer the capability to weaker models, though sharing Opus-derived Python helper code does sharply improve mid-tier agents. The study reveals capability stratification that mainstream benchmarks like SWE-Bench Verified compress into narrow bands, suggesting frontier agents succeed by constructing and debugging working models of unfamiliar environments rather than pattern-matching to training data.
Function-Aware Fill-in-the-Middle Mid-Training Improves Coding Agent Foundation Models
Researchers propose a self-supervised mid-training objective called function-aware fill-in-the-middle (FIM) that exploits the structural isomorphism between a coding agent's action-observation-continuation loop and function call sites in ordinary code. Applied to Qwen2.5-Coder-Instruct (7B/14B) and Qwen3-8B on a 2.6B-token GitHub corpus, the method yields +2.8 to +5.4 point gains on SWE-Bench-Verified and SWE-Bench-Lite across multiple post-training pipelines. Notably, the technique also mitigates capability erosion on non-agent coding and tool-use benchmarks, suggesting the function-call inductive bias generalizes beyond the training domain.
Probe-and-Refine Tuning improves coding agent performance via iterative repository guidance refinement
A new arXiv paper introduces probe-and-refine tuning, a procedure that uses synthetic bug-fix probes to iteratively improve AGENTS.md repository guidance files for LLM-based coding agents without requiring an agent loop during tuning. Evaluated on SWE-bench Verified with Qwen3.5-35B-A3B, the method achieves 33.0% mean resolve rate versus 28.3% for a static knowledge base baseline and 25.5% for an unguided baseline. The improvement is attributed to coverage gains—refined guidance helps agents locate the correct files rather than improving patch quality—and a step-budget experiment shows guidance is necessary for agents to productively use larger compute budgets.
MetaEvolve framework trains LLMs to self-evolve via meta-skill reinforcement learning
Researchers introduce MetaEvolve, a framework that cultivates 'meta-skills' such as self-reflection and multi-round refinement in LLMs through a data synthesis pipeline and evolution-aware reinforcement learning. Grounded in coding tasks where program execution provides continuous reward signals, the system trains on evolution trajectories containing programs, fitness scores, and attempt histories. MetaEvolve outperforms the strongest baseline by 10.01% on in-distribution coding benchmarks and 24.12% on out-of-distribution tasks, with a 46.9% relative improvement on open-ended algorithm optimization problems outside the training domain. The work is motivated by AlphaEvolve-style test-time scaling and aims to produce generalizable meta-skills transferable to domains with sparse training signals.
MOSS: Self-Evolving Agents via Source-Level Code Rewriting
MOSS is a system enabling autonomous agents to self-evolve by rewriting their own source code rather than being limited to text-mutable artifacts like prompts or skill files. The system anchors each evolution cycle to production-failure evidence, delegates code modification to an external coding-agent CLI, and verifies candidates by replaying failures in ephemeral trial workers before promoting via consent-gated container swap with rollback. On the OpenClaw benchmark, MOSS improves a four-task mean grader score from 0.25 to 0.61 in a single cycle without human intervention. The authors argue source-level adaptation is strictly more general than text-layer evolution, being Turing-complete and immune to long-context drift.
Goedel-Architect achieves state-of-the-art formal theorem proving with blueprint-based agentic framework
Goedel-Architect is an agentic framework for formal theorem proving in Lean 4 that uses blueprint generation — a dependency graph of definitions and lemmas — rather than recursive decomposition, enabling parallel lemma closure and global refinement. Built on DeepSeek-V4-Flash (284B-A13B), it achieves 99.2% pass@1 on MiniF2F-test and 75.6% on PutnamBench, scaling to 100% on MiniF2F, 88.8% on PutnamBench, and 4/6 on IMO 2025 when seeded with natural-language proofs. The authors claim state-of-the-art performance for an open-source pipeline at up to 500x lower cost than comparable systems.
Mistral releases Leanstral 1.5, a state-of-the-art open-source formal verification model
Mistral AI released Leanstral 1.5, an Apache-2.0 licensed mixture-of-experts model with 119B total and 6B active parameters, specialized for formal verification in Lean 4. The model saturates miniF2F (100%), solves 587/672 PutnamBench problems, and achieves new state-of-the-art results on FATE-H (87%) and FATE-X (34%), while costing roughly $4 per problem versus ~$300 for comparable systems. Trained via mid-training, supervised fine-tuning, and reinforcement learning with CISPO, it demonstrates strong test-time scaling and practical code verification capabilities, uncovering 5 previously unknown bugs across 57 open-source repositories.