AI Papers of the Week
Every paper worth reading in AI, hand-picked one week at a time.
Harness-Level Forgetting
Continual learning has always tracked what changes in the weights. Modern agents accumulate their experience somewhere else entirely, across prompts, memories, tools, skills, and routing rules, and nobody has been measuring what happens when that layer drifts.
Strategy Lock-In
Agents post-training other agents is one of the more load-bearing assumptions in current recursive self-improvement arguments. This paper analyzes a large corpus of publicly released post-training trajectories to see whether the loop actually closes, and finds a specific structural failure.
SocialRL
The dispositions that make an assistant pleasant make it a poor delegate. A friendly frontier model volunteers its principal's private information and concedes at the first sign of resistance, which is exactly the wrong behavior when it is negotiating on your behalf.
Stealing Reasoning Traces
Frontier providers hide chain-of-thought and hand the client an encrypted block instead, which the client returns with every subsequent request. This work identifies an architectural flaw in that design and turns it into a scalable extraction attack across three providers.
Reason Wide, Not Deep
Reasoning modes beat non-reasoning modes on multi-step agentic tasks and charge a 3x to 6x output-token premium on every single episode. Much of that spend goes into re-deriving procedures the model already worked out on earlier episodes in the same domain, which means the cost is recurring by accident rather than by necessity.
CEDAR
Complex systems research models feedback-driven phenomena from population dynamics to economic policy, and its central open problem is that nobody can predict how feedback structure gives rise to emergent behavior, which makes goal-directed design very hard. CEDAR, from Sakana AI, attacks that with LLM agents running Monte Carlo Tree Search over the space of feedback structures rather than tuning parameters on a fixed one. Systems are represented as a restricted runnable subset of Python with domain-specific primitives so the models can edit dynamics directly, an LLM Judge scores emergent behavior against the stated goal as a fitness function, and an LLM Editor proposes variants as a variation operator. The formalization is an MCTS variant with an LLM-parameterized transition kernel and value function, which preserves solution diversity while searching, and the LLM-based interpretability makes it possible to read back how a structural change produced the behavior.
Sample More Reflect Less
Methods that make a model criticize and rewrite its own answer nearly all generate far more text than a single chain of thought. Since generating more text raises accuracy on its own, a reported gain leaves open whether the method's idea is what helped. This paper reruns the comparison as a designed experiment.
Prompt-Induced Waste
Two prompts can request the same code change and produce the same correct patch while causing a coding agent to perform radically different kinds and amounts of work. This preregistered study measures that effect across 4,644 valid runs, 24 deterministic coding tasks, seven reasoning models, and two real harnesses.
ReOPD
On-policy distillation for agentic tasks is expensive because every update needs fresh student rollouts through the environment plus teacher queries at each visited history. Microsoft Research and the University of Amsterdam propose reusing pre-collected teacher trajectories instead.
Invisible Reasoning
Chain-of-thought monitoring rests on the assumption that a model expresses its reasoning in its output tokens. This work demonstrates a concrete failure of that assumption in models shipping today.
Global Workspace in LLMs
This Anthropic interpretability work gives a mechanistic account of when a model's verbalized reasoning is load-bearing and when it is not. It identifies a small, privileged set of internal representations that behaves like the global workspace some neuroscientists tie to conscious access.
Metacognition in LLMs
Confidence calibration, self-verification, knowing when to stop, and knowing what you do not know have mostly been studied in isolation. This survey from Yale and UC Irvine argues they are facets of one capability, metacognition, and organizes the field around a monitor and control loop wrapped around the language model.
Harness Evolution, Rethought
Automatic harness evolution is what many teams now use to squeeze more out of agents, but the reported gains might not be coming from the harness at all. This paper argues that harness evolution is itself a search procedure and must be compared against simple search baselines under matched budgets.
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.
Always-On Agents
Always-on agents are systems whose future behavior depends on durable state built up across earlier interactions, and this 130-plus page survey argues that state is far more than memory. It spans task ledgers, permissions, credentials, commitments, provenance, triggers, and effects the agent has already committed to the outside world.
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.
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.
Reasoning Quality Emerges Early
Curating reasoning data is expensive because scoring a trace usually means reading it to the end, but this UCLA work shows the quality of a trace is largely decided in its opening tokens. A short prefix predicts whole-trace quality well enough to rank and filter on, and difficulty can be detected from the loss of the first 100 tokens at a perturbed checkpoint. That turns curation into a cheap early-stopping problem, outperforming baselines while being far more token efficient at building SFT data for reasoning models.
Sakana Fugu
Frontier LLMs keep advancing, and different providers are increasingly specializing in distinct domains, which raises a natural next objective: how do you combine those individual specializations into one collectively intelligent system? Sakana Fugu answers with a family of orchestrator models that are themselves language models trained to read a user query and dynamically devise the agentic scaffold needed to solve it.
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.
Critique of the Agent Model
The word agent now covers everything from a for-loop with tool calls to speculative machine superintelligence, which makes it nearly useless as a technical term. This position paper from Eric Xing and collaborators tries to fix that by asking what an agent actually is and what agency consists of, drawing on Descartes and on science-fiction portrayals of autonomous beings to ground the discussion.
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.