AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Failure as a Process
When a coding agent fails a task, the final pass or fail label hides when the run actually went wrong. This large-scale study treats failure as a timeline and annotates over 63,000 execution steps to see how coding-agent runs break down.
Less Is More for Monitors
Chain-of-thought monitoring is treated as a reliable safety layer, but this study shows monitors often detect sabotage better with less access, not more. Giving a monitor the full agent reasoning trace can bury the signal, and the authors introduce extract-and-evaluate monitoring, where one monitor isolates the relevant excerpts and another scores them, lifting sabotage detection by 16.8 points on BigCodeBench-Sabotage with a GPT-4.1-mini monitor. Agents unaware they are watched are far easier to catch, and monitors reading long traces gain the most from filtering.
GFlowRL
Reward-maximizing RL tends to collapse large reasoning models onto a single dominant mode, and GFlowNet-style training is appealing because it matches reward distributions and keeps diverse reasoning paths. GFlowRL scales this to modern post-training by replacing the hard-to-learn partition function with an in-batch Monte Carlo estimate computed from the rollout group the pipeline already produces. It is the first GFlowNet-style RL algorithm to train stably across both dense and sparse architectures, reaching a 2048 Codeforces rating at 14B and outperforming prior methods on math, code, and adversarial red-teaming benchmarks like AdvBench and HarmBench.
Verification as a Scaling Axis
Verification is emerging as a distinct scaling axis alongside pre-training and test-time compute, and this Stanford, NVIDIA, and UC Berkeley collaboration builds a training-free verifier that reads a continuous, calibrated score straight off the scoring-token logits instead of trusting a discrete pass or fail grade.
HOLA
Linear-attention and state-space models compress an entire prefix into a fixed-size state, buying constant memory but overwriting earlier facts when many key-value associations compete. HOLA gives linear attention a hippocampal complement, pairing a compressive recurrent state with a small exact memory to recover long-range recall.
Puzzle-75B
Bigger mixture-of-experts models keep winning on quality, but serving them at interactive latency is still hard. NVIDIA compresses the hybrid MoE Nemotron-3-Super into Puzzle-75B-A9B and roughly doubles interactive server throughput while holding quality.
The Harness Effect
As orchestration harnesses mediate every model call, this study asks how much the harness alone moves cost and performance. It ran 22 evaluation tasks across six foundation models, then changed only the orchestration layer while holding the models constant.
ReContext
Models now support 128K context windows yet still fail to use evidence already sitting in the prompt. ReContext is a training-free inference harness for long-context reasoning that uses model-internal relevance signals to build a query-conditioned evidence pool, then replays it right before final generation while preserving the full original context.
Agent Limitations Taxonomy
Benchmark scores keep climbing, yet the same agent failures resurface across otherwise unrelated evaluations, hidden behind the leaderboard. This University of Oxford work synthesizes 27 benchmark, taxonomy, and audit papers spanning 19 benchmarks into the first cross-cutting taxonomy of LLM-agent limitations.
BlockSearch
BlockSearch runs the first systematic study of in-context retrieval at the scales real retrievers actually face, million-token corpora and length generalization far beyond training size. It introduces a 0.6B language-model retriever whose architectural and training changes improve over prior LM baselines and length-generalize up to 10 times beyond their training length, pointing toward retrievers that stay reliable as context windows keep growing.
RLVR Meets Human Likeness
RL with verifiable rewards only optimizes what you can objectively score, so style, structure, and diversity quietly collapse and reward hacking creeps in. This MIT work adds an adversarial discriminator trained on human demonstrations as a learned proxy for the human output distribution, and the generator maximizes both task accuracy and that human-likeness signal. Across bug fixing, story generation, and a reward-hacking benchmark, it preserves RLVR's accuracy gains while restoring the fuzzy properties it usually destroys, with misbehavior nearly disappearing.
Replicating ML Papers with Agents
This work tests whether a coding agent can replicate a scientific ML paper from its materials alone, using a skill that turns each paper claim into a target with recorded evidence and gating completion on workspace evidence rather than the agent's final message. Across twelve runs over four papers, all twelve workspaces pass the completion gate and all 158 recorded targets are matched with report coverage. Yet repeated runs still differ in how papers are split into targets and in numerical fidelity, so completion becomes reproducible even when the path is not.
Red Queen Gödel Machine
Self-improving agents are only as strong as the evaluator scoring them, and most systems freeze that evaluator in place, so improvement stalls the moment the judge stops getting harder. The Red Queen Gödel Machine makes the evaluator part of the search itself, letting agents and the criteria that judge them co-evolve. --- ---
The Verification Horizon
Reinforcement learning for coding agents lives or dies on the reward signal, and this Qwen work argues there is no silver bullet. As policy capability grows, any fixed reward function eventually gets gamed, so verification has to co-evolve with the generator it scores. ---
Paper Assistant Tool
AI is accelerating how fast papers get written, but peer review is still bottlenecked on human throughput, with combined submissions to the big ML conferences projected to top 73,000 this year. Google’s Paper Assistant Tool is an agentic framework built to do deep scientific review and verification at that scale. ---
Generative Skill Composition
Coding agents accumulate large skill libraries, and picking the right skills for a task has become the bottleneck. The usual options either dump the whole collection into context or retrieve skills with embeddings and rerankers, and both treat selection as a ranking problem rather than a joint plan. ---
HORIZON
HORIZON treats hardware design as repository-level code evolution, compiling a Markdown harness into a project pack with domain knowledge, an executable evaluator, an acceptance predicate, and a git and runtime policy. A hands-free agent loop then evolves an isolated git worktree, using repository operations for state management, tracing, and replay. Across ChipBench, RTLLM, Verilog-Eval, and nine CVDP categories it reaches full benchmark completion with a completely hands-free loop, extending repository-scale self-evolution from EDA software to hardware artifacts themselves. ---
Agent-Native Memory
Memory for LLM agents has quietly grown from a retrieval add-on into a full data system, with persistent storage, retrieval, update, consolidation, and lifecycle governance running throughout an agent's execution. Yet most evaluations still score memory only through end-to-end task metrics like F1 and BLEU, treating the whole stack as a black box. This paper studies agent memory from a data management perspective and asks what we are actually missing when we measure it that way.
Autodata
Building synthetic training data has mostly stayed a fixed pipeline that you hand-tune once and then freeze. Autodata rethinks that by casting an AI agent as a data scientist that builds high-quality training and evaluation data, then meta-optimizes that agent so it learns to create even stronger data over time.
Agent-as-a-Router
Most users now have access to many LLMs that each excel in different domains, so routing each task to the right model matters for both quality and cost. Existing routers treat this as a static, one-off classification problem, and this paper shows that framing is exactly what holds them back.
Skill-MAS
Automatic generation of multi-agent systems is stuck between inference-time methods that reuse frozen frontier models but never learn, and training-time methods that internalize experience through gradient updates but are capped by the weaker models small enough to fine-tune. Skill-MAS proposes a third path that treats high-level orchestration as an evolvable Meta-Skill, decoupling experience retention from weight updates so frontier models keep getting better at orchestration without any gradient steps. Across four complex benchmarks and four distinct LLMs it delivers strong, transferable gains at a favorable cost-performance trade-off.
Reliability without Validity
LLM-as-a-Judge is the default way to evaluate language models, but validating those judges with exact-match agreement never corrects for chance and systematically overstates how good they are. In the largest audit to date, spanning 21 judges from nine providers across MT-Bench, JudgeBench, and RewardBench over 118 runs and roughly 541,000 judgments, the gap between raw agreement and chance-corrected Cohen's kappa runs 33 to 41 percentage points, rankings shift by up to 14 positions across benchmarks, and high test-retest reliability coexists with severe position bias. The authors distill their findings into a Minimum Viable Validation Protocol so teams can stress-test judges before trusting them.
NatureBench
Can coding agents move past reproduction toward actual discovery on real scientific problems? NatureBench distills 90 cross-discipline tasks from peer-reviewed Nature-family papers and runs them in NatureGym, an automated pipeline that builds a standardized containerized environment per task to fix the environment-fragmentation problem. Under a strict web-search-disabled protocol, the strongest of ten frontier agent configurations beats published SOTA on only 17.8% of tasks, and analysis shows agents win mainly by translating problems into familiar supervised prediction rather than through genuine scientific invention.
SpatialClaw
Spatial reasoning over 3D and 4D scenes is still where general vision-language models break down, because they emit a text answer directly rather than measuring anything. From NVIDIA, SpatialClaw is a training-free framework that rethinks the action interface and lets a VLM-backed agent reason through code instead. The agent writes one Python cell per step into a persistent Jupyter kernel preloaded with perception primitives and scientific libraries, then inspects intermediate results and revises its strategy across steps.