AI Coding Live News

HarnessPro (arXiv 2610.26770, submitted Oct 19 and revised Oct 20), from Tsinghua University, Z.AI (Zhipu) and collaborators, is a label-free way to improve an agent harness: specialized agents diagnose execution traces, rewrite instructions and runtime policies, score candidates with LLM-generated tests, and keep Lite, Medium and Heavy harnesses along the performance-cost frontier. Per the paper, with GPT-5.5 and Codex the Heavy mode reaches 80.8% on Terminal-Bench 2, 4.5 points above the prior RHO method, while cost per task drops from $1.09 to $0.72. The code is MIT-licensed and drew wider media coverage this week, including from Pandaily.

Zoom

Tencent ARC in collaboration with Tsinghua University released and open-sourced Looking Ahead (arXiv:2610.29989, GitHub: TencentARC/Looking-Ahead), establishing a breakthrough paradigm for autoregressive video synthesis. Overcoming the chronic pathology of traditional autoregressive video models—where unidirectional sequential extrapolation causes exponential error accumulation and severe visual collapse over extended rollouts—Looking Ahead introduces Lookahead Attention. By coupling casual autoregressive decoding with a forward-looking planning stream that anchors future spatiotemporal keypoints, the framework eliminates cumulative drift and visual jitter in long-form generation.

Zoom

Researchers from Tsinghua University and CUHK introduced a breakthrough paper titled 'Cache Engineering for Agents' (arXiv:2610.28577) alongside the open-source AKV serving engine (thu-nics/AKV). While multi-turn agents rely on traditional 'context engineering' (text trimming and summarization) that invalidates prefix KV-caches and incurs punishing re-prefilling overhead, cache engineering empowers agents to directly manipulate KV-cache states via 'drop' and 'repos' primitives. Coupled with AKV's drop-aware radix-tree engine, the framework slashes agent inference costs by 27.7%, accelerates P95 latency by 1.5x, and boosts benchmark solve scores by 1.7 points.

Zoom

Current memory benchmarks for video world models conflate whether an action was carried out with whether the model remembered its outcome, incorrectly attributing action execution failures to cognitive memory loss. Researchers from HKU and Tsinghua introduced MIST (arXiv:2610.23542, 12 upvotes on Hugging Face). By decoupling physical execution from temporal memory, MIST reveals that up to 70% of apparent memory failures in leading models are actually execution failures. Correcting for execution raises true scene memory accuracy by +11.8 to +29.6 percentage points.

Zoom

Researchers from Tsinghua University open-sourced Dynamic Band Splitting (DBS, arXiv:2610.14993, 134 upvotes on Hugging Face Papers), replacing fixed predefined frequency partitions in band-split audio source separation models with content-conditioned Top-K boundary selection and adaptive pooling under a fixed token budget. Integrating DBS into BSRNN yields a +29.0% relative improvement in average SI-SDR on the DnR v2 benchmark.

Zoom

Addressing the representation gap between high-level VLM semantics, video world models, and low-level control in Vision-Language-Action (VLA) architectures, PINT (arXiv:2610.16355) introduces learnable chunk-level Predictive Intent Tokens to reach 15.5% on RoboTwin 2.0-Hard without embodied pretraining (4.8x over StarVLA's 3.2%). Alongside PAIR (arXiv:2610.09016) and PKU's SimForcing (arXiv:2610.06598), these frameworks bridge perception and action with zero inference overhead.

Zoom

Vision-language navigation requires abundant, high-quality language supervision, yet traditional synthetic pipelines generate supervision from static models that lag behind evolving navigation policies. Shanghai AI Lab, Tsinghua, and SJTU introduce RSI-Nav, an embodied recursive self-improvement framework combining instruction-following and instruction-generation in a single model. Using shard-wise dual bootstrapping on unlabeled trajectories, RSI-Nav achieves state-of-the-art 68.6% SR on R2R-CE and 72.4% SR on RxR-CE without human intervention.

Zoom

Token-level LLM routing—routing easy tokens to small models and switching to frontier models for critical reasoning tokens—presents an optimal trade-off between output quality and inference cost. However, current LLM serving engines (such as vLLM and TGI) are architected for single-model execution, suffering severe step desynchronization, batch admission latency, and extreme orchestration complexity when handling high-frequency token transfers across heterogeneous models. Researchers from Tsinghua University's NICS Lab present TokenRouter (arXiv:2610.12242, accepted at NeurIPS 2026, code: github.com/thu-nics/TokenRouter), an efficient and developer-friendly serving system engineered for token-level routed LLM inference. TokenRouter introduces a 'request-centric programming and model-centric execution' paradigm, enabling developers to write routing logic intuitively per request while the runtime handles asynchronous execution across subservers. Incorporating a Delayed-Batching Scheduler parameterized via discrete-time Markov chain modeling, a decoupled tri-loop architecture, and an instant handoff-resume mechanism, TokenRouter achieves 2.01x to 64.15x higher decoding throughput compared to existing serving platforms across diverse routing workloads, with complete code open-sourced.

Zoom

Terminal coding agents are expanding from standard software engineering into scientific domains like bioinformatics, materials science, and physics. However, building training sandboxes requires executable reference behavior and rigorous verifiers capable of distinguishing true semantic correctness from superficial syntactic mimicry—a process that currently requires tedious manual engineering per task. Researchers from Tsinghua University and Tencent introduce Software-in-the-Loop Reconstruction (SWR, arXiv:2610.02710), a self-supervised environment scaling framework. SWR extracts ground-truth outputs and verifiers from existing executable scientific software workflows by perturbing input configurations into public observations and hidden evaluations. Agents synthesize complete, editable programs from input-output schemas without viewing source code, evaluated by hierarchical verifiers testing semantic fidelity, structural integrity, and anti-shortcut compliance. Implemented across 500 workflows and 46 software suites spanning six scientific domains, SWR generates 1,422 verified trajectories. Supervised fine-tuning of Qwen3.8-27B elevates Terminal-Bench 2 average success from 47.94% to 53.56%, establishing a scalable paradigm for scientific terminal agent training.

Zoom

As autonomous agents execute increasingly long-horizon workflows, traditional strategies relying on raw message history or lossy memory compression fail to provide coherent comprehension of evolving environment states. Agents frequently succumb to 'Belief Trapping'—iteratively generating uncoordinated tool calls without moving toward the goal. Researchers from Tsinghua University and Nankai University introduce PoS (Possibility-of-State), an inference-time framework that constructs and dynamically maintains explicit belief states. Coupling current world-state estimates with unresolved task requirements, PoS continuously verifies belief consistency, detects belief traps in real time, and triggers targeted recovery protocols. Across four execution and system diagnosis benchmarks on three frontier LLM backbones, PoS achieves top performance with robust resilience against context explosion.

Zoom

Pretrained transformers utilize remarkably little of their internal depth when following contextual reference chains: thirteen base models reliably follow only 1.4 to 3.6 lines, with pretrained loops offering minimal relief. Researchers from Tsinghua and Peking University discover that transformers 'stop thinking too early' due to early computation stalling. By inserting a tiny, task-trained rank-8 LoRA at just one single early layer while keeping all original parameters completely frozen, they initiate a computational 'relay' across middle layers. On 24-line reference chains, Qwen3-8B's exact accuracy leaps from 15.5% to 99%, with extended training scaling past 50 to 160 lines and boosting multi-hop reasoning on MuSiQue.

Zoom

Reinforcement learning for code agents typically relies on executable tests for binary rewards, leading algorithms like GRPO to assign identical scalar advantages to all passing rollouts regardless of code bloat, extraneous modifications, or architectural hygiene. Researchers from ByteDance, Peking University, and Tsinghua introduced GAGAR, a quality-aware advantage redistribution framework. By staging rollouts in a shared workspace where an SFT agentic grader ranks passing candidates and dynamically shifts credit toward clean solutions, GAGAR successfully stabilizes code agent RL on 310B Flash and 1.02T Pro models.

Zoom