Handoff Tax bills trajectories, Open-MOPD hits 83%, SWE agents finish 5%

Sources

The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents huggingface.co

Coding agents perform long-running tasks spanning dozens of model calls, tool uses, and code edits. As these runs unfold, users face a practical cost-quality trade-off: escalating to a stronger model when a cheaper one struggles, or downshifting once the hard reasoning is complete. Each switch requires the receiver to continue a non-native trajectory produced by another model. We study how this handoff affects quality and cost, and how varying the trajectory information inherited by the receiver

Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation huggingface.co

Multi-teacher on-policy distillation suffers from token-level budget misallocation across domains, which is addressed by balancing, dynamic allocation, and reward refresh to recover most of the oracle ensemble’s capability.

D^3-MOPD: Adaptive Dynamic Domain ScheDuling for Efficient Multi-Teacher Distillation huggingface.co

D³-MOPD dynamically adjusts domain sampling ratios during multi-teacher distillation by monitoring per-domain reverse-KL trajectories, improving convergence efficiency and closing most of the student-to-teacher performance gap.

SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration? huggingface.co

The study introduces a benchmark for evaluating autonomous software migration by coding agents, finding that current models rarely complete migrations correctly.

Prefix Sliding for efficient test-time scaling huggingface.co

Prefix Sliding trims KV-cache costs during test-time scaling by discarding intermediate tokens deemed unimportant to later reasoning steps. The method pairs full attention with a sliding window and needs no retraining, letting existing reasoning models stretch to longer chains without extra GPU memory.

Gated Recurrent Transformers: Expressive Depth through Recurrent Modulation huggingface.co

A shared transformer block is reused across depth under adaptive update gates, delivering quality on par with or better than deeper baselines. Under isoFLOPS and isoPARAMS tests, the recurrent design cuts parameter count and memory while preserving expressivity through learned depth reuse.

JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution huggingface.co

JIT-Agent trains a model to generate composable four-module harnesses tailored to each task, boosting off-the-shelf LLMs across DeepSearchQA, OdysseyBench and OpenCode. The approach self-evolves via distillation and lifts performance on models from Claude Code down to smaller open weights.

Is Next-Chunk Reasoning RL Really Better than SFT? Revisiting Training Strategies under no-CoT Data huggingface.co

Revisiting post-training under no-CoT corpora, mixed supervised fine-tuning matches or exceeds next-chunk RLVR on math and out-of-domain accuracy while running more efficiently. The result challenges the default assumption that reinforcement learning is the stronger recipe when chain-of-thought traces are unavailable.

Code World Model: Coding Agent as World Brain huggingface.co

Code World Model splits simulation into two parts: a language model emits executable code that updates persistent world state, while a video model renders observations from proxy representations. The decomposition enforces spatiotemporal consistency that pure video world models routinely violate.

LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale huggingface.co

LibriBrain100 offers 100 hours of magnetoencephalography recordings aimed at neural speech decoding at scale. Experiments show that deep within-subject data plus multi-subject supervised fine-tuning of pre-trained models sharply improves decoding accuracy, giving the brain-computer interface field a much larger training substrate.

Skill Issue: Are Skills Language-Invariant in LLMs? huggingface.co

Pitting LLMs against themselves in text-based games across languages exposes large cross-lingual gaps in spatial reasoning, imperfect-information play and resource allocation. Switching the intermediate reasoning language recovers part of the deficit, suggesting skills are tied to language rather than being fully invariant abstractions.

FrontierChallenge: Evaluating Scientific Workflow Completion huggingface.co

FrontierChallenge evaluates end-to-end scientific workflows across domains, revealing that frontier models complete only about 20% of tasks despite high partial scores and frequent claims of completion.

Pushing the Limits of High-Resolution Weather Forecasting through Data Scaling huggingface.co

BaguanHR improves high-resolution weather forecasting by using variable-wise super-resolution to synthesize training data, overcoming the limits of coarse-resolution model transfer and demonstrating strong scaling benefits.

VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction huggingface.co

VoiceMem introduces a dual-brain streaming memory architecture for speech language models that improves retrieval accuracy, emotional personalization, and real-time efficiency.

WarpSAC: Towards the Pinnacle of Scalable Off-policy RL by Rethinking Exploration and Exploitation huggingface.co

Off-policy reinforcement learning stabilizers vary with data availability, motivating regime-aware algorithms that adapt normalization and Q-function clipping to improve efficiency across CPU and GPU-parallel training.

A Programming Paradigm for Spatiotemporal Composability huggingface.co

A calculus and framework unify revertible effects and reactive coeffects into a context paradigm enabling spatiotemporal composability for dynamic software components.

GUI-Primitives: Diagnosing Spatial Reasoning Failures in Vision-Language GUI Grounding huggingface.co

A benchmark of contrastive GUI instructions reveals that vision-language models mostly fail to localize interface elements rather than misunderstand spatial relations, though marking candidates substantially improves selection.

Agent-G^2: Gaussian Guidance for Agentic Reinforcement Learning huggingface.co

Agent-G² models hint depth as a Gaussian distribution estimated online from existing rollouts, improving reinforcement learning on long-horizon tasks without extra probing.

Long-Horizon Audio-Visual Generation for Persistent Stories and Interactive Worlds huggingface.co

JoyAI-Echo-1.5 unifies long-form video and interactive world generation through cross-shot memory, geometry-aware camera control, and rollout-aware training to maintain identity and coherence over extended sequences.

Stream4D: 4D-Consistency for Streaming Autoregressive Diffusion Video Models huggingface.co

Stream4D improves autoregressive video generation by replacing static 3D critics with a dynamic 4D reconstruction reward and motion prior to preserve coherent motion and reduce geometric drift.

FIRM-Video: Check Before You Score for Reliable Text-to-Video Reward Modeling huggingface.co

FIRM-Video uses checklist-driven verification of temporal visual evidence to build reliable reward models for text-to-video evaluation and alignment.

V-Rubrics: Visual Faithfulness via Rubric-Based Reinforcement Learning huggingface.co

Visual Rubrics-Based Reinforcement Learning improves vision-language model grounding by scoring answers on visual faithfulness, reasoning consistency, and instruction following using structured partial credit.

Are Android GUI Agents Robust Against Runtime Anomalies? AnTrap: Evaluating Agents in Dynamic Adversarial Environments huggingface.co

AnTrap benchmarks GUI agent robustness by injecting dynamic anomalies into execution trajectories, revealing universal vulnerabilities and distinguishing learnable traps from intrinsic reasoning limits.

StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models huggingface.co

StreamPI enhances single-frame vision-language-action models with streaming temporal reasoning via instruction-anchored attention and randomized interval training, improving robot manipulation without extra parameters.

MA-VLA: Multi-Arm Vision-Language-Action Model for Collaboration and Compositional Generalization huggingface.co

MA-VLA enables multi-arm collaboration by assigning atomic actions to individual arms and using training-time permutations to generalize to unseen coordination patterns.

RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval huggingface.co

RetrievalRouter adaptively selects retrieval pipelines per query to improve both accuracy and speed across diverse document benchmarks.

Rubrics as Visual-Repair Context for Self-Evolving UI-to-Code Generation huggingface.co

RubSE improves UI-to-code generation stability by using rubric-guided self-evolution to prevent visual repair coupling and trajectory collapse.

VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning huggingface.co

VBBR-Pro introduces a closed-loop testbed that enables scalable, verifiable, and controllable native visual reasoning through generation across diverse visual substrates.

VGI-Bench: Probing Visual Intelligence in Video Generation Models huggingface.co

VGI-bench evaluates visual reasoning in video generation models through 27 tasks, revealing limited reliability and self-correction capabilities.

Video-IFBench: Evaluating Instruction Following of Multimodal LLMs in Video Understanding Scenarios huggingface.co

A new benchmark evaluates how well multimodal language models follow diverse video-based instructions with visual, audio, and structural constraints.

Super Star: Towards Streaming Real-time Interactive Agents for Digital Humans huggingface.co

A real-time framework for online co-speech gesture generation uses a causal multimodal autoregressive model with streaming speech and motion history, supported by synthetic dialogue data and continual user-feedback adaptation.

A Modular Agent for Reliable and Auditable Spatial Relation Verification in CT Scans huggingface.co

A modular medical imaging agent decomposes spatial relation verification into parsing, anatomical localization, and geometric rules to outperform end-to-end vision-language models on CT spatial reasoning.

Real-TurnTurk: A Multimodal Turkish Corpus for Turn-Taking Prediction huggingface.co

A multimodal Turkish dialogue dataset and genetic algorithm-optimized interpretable rules are used to predict turn transitions from visual, acoustic, and linguistic cues.

References

SSOJet — cheapest AI coding models ssojet.com

High-end models like Claude 4.5 Opus lead in accuracy (80.9%) but can cost up to $0.68 per resolved task… DeepSeek V3.2 reportedly costs as little as $0.028 per task, a 24x difference.

beam.ai — ICLR 2026 coverage of Lost-in-Conversation beam.ai

State-of-the-art models experience an average accuracy drop of 39% when subjected to these ‘sharded’ instructions… if these shards are simply concatenated into a single prompt at the end, performance recovers to nearly 95% of the baseline.

Entelligence — Loop Engineering Explained entelligence.ai

Tools like Devin allow users to ‘scrub the timeline’ and restore a specific checkpoint… If a full restart is necessary, the recommended strategy is to restart with a tighter brief—a condensed instruction list that includes the previous attempt’s failures.

alphaXiv review of arXiv:2608.24358 alphaxiv.org

Treating context as ‘neutral state’ is a fundamental design error in current agent orchestration.

JetBrains Research — Efficient Context Management (Dec 2025) blog.jetbrains.com

‘Lost-in-the-middle’ effects can cause a 30% drop in accuracy when vital information is buried deep in a long, unsummarized trajectory… maintaining a compact summary of successful trajectories is more critical for reliability than simply increasing context window size.

Blackmatter VC — Loop Engineering practitioner verdict blackmatter.vc

The accumulation of context in long-running loops often accounts for over 99% of session spend… one study found agent solutions are merged by maintainers at roughly half the rate of human solutions.

Moonlight review of D³-MOPD themoonlight.io

D³-MOPD reached the same peak performance with a 3× reduction in rollout steps and closed approximately 97% of the performance gap between student and teachers, with the student actually surpassing specialist teachers on three of seven evaluated domains on Qwen3-30B-A3B.

Thinking Machines Lab — On-Policy Distillation blog thinkingmachines.ai

On-policy distillation samples trajectories from the student and uses a teacher to score every token with dense reverse-KL supervision; the objective can be dropped into an RL loop (GRPO/PPO) with roughly ten lines of PyTorch.

ByteDance Seed announcement — SIA Lab (Doubao Seed × Tsinghua AIR) seed.bytedance.com

The Scalable Intelligent Model Technology Joint Research Center (SIA Lab), established in late 2024 by ByteDance’s Doubao Seed Team and Tsinghua AIR, is the institutional home for the MOPD/Open-MOPD/D³-MOPD line of work targeting the MiMo and Doubao model families.

GitHub — BytedTsinghua-SIA/Open-MOPD github.com

The repository releases the full pipeline — MixSFT, three domain RL teachers, and the multi-teacher distillation stage — with local shell launchers, SmolLM3-3B checkpoints, training trajectories and evaluation data intended for reproducibility on an 8×A100-80GB node.

AlphaXiv discussion of on-policy distillation limits alphaxiv.org

A 2026 line of critique argues OPD’s gains may be partly illusory: it primarily improves sampling efficiency — making the model more likely to surface a solution it already ‘knew’ — rather than expanding the solvability boundary, which remains bounded by the teacher.

Medium analysis — Behavior Leverage Imbalance in MTKD joansantoso.medium.com

Aggregate token weighting masks extreme divergences at ‘mode-entry’ tokens like tags; Soft Clamp reduced tool-call errors from 13.7% to 9.0% on APIGen-MT, suggesting where teacher signals act matters more than their aggregate magnitude.

Remio.ai analysis of Bun Zig→Rust migration remio.ai

Andrew Kelley, the creator of Zig, criticized the resulting codebase as ‘unreviewed slop,’ highlighting that it contained over 13,000 unsafe Rust blocks—roughly 4% of the entire repository… while the migration passed its primary test suite, 19 regressions escaped into the production merge

Vals AI Code Migration leaderboard vals.ai

Claude Opus 5 at the top with a score of 57.47%, followed closely by Claude Fable 5 (55.06%) and GPT-5.6 Sol (52.92%) […] percentage of hidden behavior tests passed after a migration

Liner review of SWE-EVO liner.com

SWE-EVO introduces the Fix Rate, a metric designed to capture ‘partial progress’ on long-horizon tasks while strictly penalizing regressions […] SWE Refactor Bench uses a more binary, hurdle-based Composite Score. If a model fails the initial ‘Audit’, it receives a zero regardless of test results

Vector Labs practitioner brief vector-labs.ai

agents are effective at modernizing narrow ‘slices’ of test suites (e.g., migrating from unittest to pytest), they often leave behind ‘dead dependencies’ or unmigrated modules that create new categories of technical debt

Moonlight review of SWE Refactor Bench themoonlight.io

A primary methodology critique centers on the benchmark’s small task count of just 20 migrations […] 13 of the 20 tasks received zero accepted solutions across 520 total runs, suggesting that for the hardest 65% of the corpus, current frontier models provide no utility

Sam Codeman blog: ‘Only 28 of 520 agent runs completed a migration’ samcodeman.com

68% of submissions that passed both the migration audit and 130,000+ fixed tests were ultimately debunked by agentic verifiers, proving that static verification is insufficient for architectural migrations