Router cap is co-failure, Qwen flips coding verify, tool-RL blames formatting

Sources

When Does Combining Language Models Help? A Co-Failure Ceiling on Routing, Voting, and Mixture-of-Agents Across 67 Frontier Models huggingface.co

Multi-model systems face fundamental accuracy limits determined by the rate at which all models fail simultaneously, regardless of their individual correlations or ensemble strategies.

The Verification Horizon: No Silver Bullet for Coding Agent Rewards huggingface.co

Verification challenges in AI agents arise from the difficulty of aligning proxy signals with human intent, requiring adaptive verification systems that evolve alongside generative capabilities.

Why Multi-Step Tool-Use Reinforcement Learning Collapses and How Supervisory Signals Fix It huggingface.co

Research investigates how different supervisory signals and training strategies improve the stability and performance of large language models in tool-use tasks, addressing issues like catastrophic collapse and format sensitivity through interleaved supervised fine-tuning and reinforcement learning.

Running the Gauntlet: Re-evaluating the Capabilities of Agents Beyond Familiar Environments huggingface.co

A new web-based benchmark called Gauntlet stress-tests agents outside familiar environments and finds wide gaps versus humans on temporal perception, graphical understanding, and 3D reasoning. The automated evaluation engine runs a modular pipeline of vision-intensive tasks where current agentic systems consistently underperform.

Neglected Free Lunch from Post-training: Progress Advantage for LLM Agents huggingface.co

Reinforcement learning post-training already produces an implicit step-level scoring signal, derived as a log-probability ratio the authors call progress advantage. The trick removes the need to train a dedicated reward model and supports test-time scaling, uncertainty quantification, and failure attribution for LLM agents.

Hallucination in World Models is Predictable and Preventable huggingface.co

Hallucinations in learned world models cluster in low-data regions of state-action space, making them predictable from data-centric signals. Coverage-aware sampling and curiosity rewards steer fine-tuning toward those gaps, cutting visual world-modeling errors without needing ground-truth actions or rewards.

OpenBioRQ: Unsolved Biomedical Research Questions for Agents huggingface.co

OpenBioRQ evaluates agentic models on open biomedical research questions where correctness hinges on source verification rather than a fixed answer key. Both frontier and open-weight agents show retrieval-grounded reasoning failures and tool-use collapse, frequently fabricating citations under pressure.

GUI vs. CLI: Execution Bottlenecks in Screen-Only and Skill-Mediated Computer-Use Agents huggingface.co

A controlled benchmark gives screen-only GUI agents and skill-mediated CLI agents identical goals, initial states, and verifiers across 440 desktop tasks spanning 18 applications and 12 workflow categories. The setup isolates interaction modality from confounds, exposing where each execution layer bottlenecks.

Discretizing Reward Models huggingface.co

Continuous reward models often assign different scores to equally good responses, destabilizing policy learning. Discretizing their outputs preserves discriminative ability while cutting oversensitivity, and pairs with Monte Carlo dropout for specificity estimates that reduce reward hacking during RL fine-tuning.

Information-Aware KV Cache Compression for Long Reasoning huggingface.co

InfoKV compresses the KV cache by scoring token importance with information-theoretic signals like predictive uncertainty and Forward Influence, not attention weights alone. The entropy-aware framework preserves long-context reasoning accuracy in LLMs while shrinking memory needed for extended chains of thought.

JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting huggingface.co

JetSpec is a speculative decoding framework that combines efficient forward drafting with causal conditioning to improve LLM inference speed and acceptance rates across various benchmarks.

How Post-Training Shapes Biological Reasoning Models huggingface.co

Post-training stages in biological reasoning models differently affect generalization, with continued pre-training aligning models with biological language, supervised fine-tuning improving in-domain performance but reducing out-of-domain generalization, and reinforcement learning recovering out-of-domain performance when applied to well-aligned checkpoints.

Confidence-Aware Tool Orchestration for Robust Video Understanding huggingface.co

Robust-TO addresses the Blind Trust Problem in video reasoning by integrating per-frame trustworthiness into an agentic framework that improves accuracy under realistic perturbations through calibrated evidence weighting and reliability-aware reasoning.

In-Context World Modeling for Robotic Control huggingface.co

ICWM enables robot policies to infer system variables from self-generated interactions, allowing adaptation to novel configurations without parameter updates by treating system identification as an in-context adaptation problem.

ViQ: Text-Aligned Visual Quantized Representations at Any Resolution huggingface.co

ViQ presents a visual quantization framework that balances semantic richness and detail preservation in discrete representations, enabling efficient multimodal training with native-resolution inputs.

ABACUS: Adapting Unified Foundation Model for Bridging Image Count Understanding and Generation huggingface.co

ABACUS is a unified vision-language model that performs object counting and related tasks through innovative spatial grounding, boundary-aware counting policies, and self-critical learning strategies.

PhysiFormer: Learning to Simulate Mechanics in World Space huggingface.co

PhysiFormer uses coordinate-space diffusion to generate physically-plausible 3D object motions without explicit inductive biases, enabling efficient multi-object reasoning and generalization to complex materials and geometries.

Fast LeWorldModel huggingface.co

Fast-LeWM accelerates visual planning by replacing autoregressive rollout with parallel action-prefix prediction, reducing computational costs and latency accumulation during long-horizon predictions.

EO-WM: A Physically Informed World Model for Probabilistic Earth Observation Forecasting huggingface.co

EO-WM is a video diffusion transformer for multispectral Earth Observation forecasting that incorporates physically informed conditioning frameworks to better capture weather-driven uncertainties in land-surface dynamics.

CoffeeBench: Benchmarking Long-Horizon LLM Agents in Heterogeneous Multi-Agent Economies huggingface.co

CoffeeBench evaluates LLM agents in a multi-agent economic simulation where firms interact over 90 days to maximize profits, revealing differences in communication patterns and performance among various models.

Qwen-Image-Agent: Bridging the Context Gap in Real-World Image Generation huggingface.co

A unified agentic framework called Qwen-Image-Agent is proposed to address the context gap in text-to-image generation by progressively constructing complete generation context through planning, reasoning, searching, and memory mechanisms.

OPID: On-Policy Skill Distillation for Agentic Reinforcement Learning huggingface.co

On-policy skill distillation framework extracts dense hindsight supervision from completed trajectories to improve language agent training efficiency and performance.

DanceOPD: On-Policy Generative Field Distillation huggingface.co

A novel on-policy generative field distillation framework called DanceOPD is proposed to unify text-to-image generation, local editing, and global editing capabilities in flow-matching models through capability-specific routing and velocity-based training.

LISA: Likelihood Score Alignment for Visual-condition Controllable Generation huggingface.co

Score-based generative modeling reveals that side networks contribute likelihood scores to conditional control, leading to improved training efficiency through likelihood score alignment regularization.

COrigami: An AI Pipeline for Co-Designing Flat-Foldable Visually Recognisable Origami huggingface.co

A computational origami system generates crease patterns from natural language using AI-driven optimization and aesthetic evaluation, enabling human-AI collaboration in mathematically constrained design.

References

Berkeley RDI blog — ‘Trustworthy Benchmarks’ rdi.berkeley.edu

Qwen3-Coder on multiple django tasks [was] locating the fixing commit/PR via logs to bypass actual coding logic

StartupFortune — ‘SWE-bench has been benchmaxxed’ startupfortune.com

OpenAI announced it would no longer evaluate models on SWE-bench Verified, citing that 59.4% of audited problems had flawed tests or were vulnerable to reward hacking behaviors like monkey-patching graders

Medium review of WebDevJudge (ICLR 2026 Oral) medium.com

human pairwise agreement reaches 84.56%, [while] the best-performing machine judge (GPT-4.1) only manages 70.34% — a 14 to 15 point gap between automated judges and human raters

Ethayarajh et al., ICML 2024 — original KTO paper proceedings.mlr.press

KTO treats model alignment as an optimization of human utility using unary signals—binary labels of ‘desirable’ or ‘undesirable’ for a given output

Sebastian Sigl — ‘LLM Judge Biases’ sebastiansigl.com

Claude 3.5 Sonnet reportedly rates its own outputs roughly 25% higher than human panels do… larger models are not inherently more robust; GPT-4o and LLaMA-3.1-70B can be more vulnerable to misleading task cues than their smaller counterparts

TheMoonlight.io review of ‘Verification Horizon’ themoonlight.io

Qwen’s ‘user-as-verifier’ approach required over 535,000 human annotations across 125,000 trajectories… lacks a ‘practical recipe’ for automating this co-evolution without human-in-the-loop bottlenecks

MarkTechPost on Princeton’s Self-MoA marktechpost.com

Self-MoA achieved a 6.6% improvement over the standard heterogeneous MoA on AlpacaEval 2.0 and an average 3.8% gain across MMLU, CRUX, and MATH — mixing different LLMs frequently lowers the average quality of the proposers.

Google Research — Science of Scaling Agent Systems research.google

A capability ceiling appears once a single-agent baseline reaches roughly 45% accuracy; beyond this, additional coordination overhead compounds with tool complexity to reduce overall success rates.

Medium — ‘I thought more AI agents always meant better results’ medium.com

Hybrid MAS required 515% more tokens to produce results that were actually inferior to a single-agent setup, achieving only 13.6 successes per 1,000 tokens compared to the single agent’s 67.7.

OpenReview — ‘The Routing Plateau’ openreview.net

On the RouterBench suite, the top 15 evaluated routers differed by less than 0.23 percentage points in accuracy — most methods converge to a narrow performance band well below the Oracle ceiling.

AI Weekly — coverage of Chen’s paper aiweekly.co

Chen shows the wrong metric caps multi-model LLM gains; pairwise error correlation cannot detect the all-wrong tail, so two systems with identical ρ can have vastly different co-failure ceilings.

arXiv — algorithmic monoculture / error correlation study arxiv.org

Larger and more accurate models tend to show even higher error correlation than their smaller predecessors — as models converge toward ground truth, they also converge on the same ‘hard’ problems.

The Moonlight review — ‘On GRPO Collapse in Search-R1: Lazy Likelihood-Displacement Death Spiral’ themoonlight.io

the model’s likelihood for correct trajectories begins to decay … low-confidence responses perversely inflate gradients from incorrect trajectories when compared to a group mean … causing entropy spikes and an abrupt drop in rewards within the first few hundred training steps

AIWeekly summary of the CASIA paper aiweekly.co

vanilla RL training can cause sudden, extreme spikes in the probability of specific control tokens, such as , , or … the model appears to have lost its reasoning ability when it has actually only lost its ability to format that reasoning

EmergentMind — RAGEN / StarPO-S ‘Echo Trap’ analysis emergentmind.com

RAGEN identifies the ‘Echo Trap’ — sudden reward variance cliffs and gradient spikes during training … a behavioral loop where agents repetitively execute actions that provide high immediate rewards but lead to suboptimal long-term reasoning

DAPO paper (arXiv 2504.11536) arxiv.org

clip-higher sets a more relaxed upper bound (ε_high > ε_low), allowing the model to more aggressively push up the probability of exploration tokens … DAPO-trained models outperformed DeepSeek-R1 variants by reaching higher accuracy in 50% fewer training steps

ReTool project page retool-rl.github.io

ReTool-32B achieved 72.5% accuracy on AIME, surpassing OpenAI’s o1-preview by nearly 28% while requiring 60% fewer training steps … a ‘cold-start’ supervised fine-tuning phase followed by RL that focuses purely on final outcomes

ThinkPRM / generative process reward model paper (arXiv 2511.06380) arxiv.org

generative PRMs such as ThinkPRM use long Chain-of-Thought reasoning to self-verify, outperforming standard verifiers using only 1% of the typical process labels