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.
The Control-Plane Tax
Serving systems for agentic applications are still built on assumptions inherited from single-turn LLM inference. This work instruments ten real agentic applications end to end and finds the model is often not what your latency bill is paying for.
Demystifying Agent Skills
Skills are usually assumed to inject knowledge the model lacks. This paper runs the controlled comparison and finds that assumption is almost entirely wrong, which changes what a good skill should contain.
The Fragility of Self-Improving Agents
Memory-based self-improving agents report gains that have never been checked against evaluation noise. This re-evaluation adds the two things prior work skipped, multiple runs to measure variance and randomly shuffled task orders, and both hurt. Agent evaluation is already noisy on multi-step tasks, and stacking a self-improvement loop on top amplifies that noise rather than averaging it out. The sharper finding is that default task orderings impose an implicit curriculum, and much of the reported gain was riding on it. Adding detailed rubrics and environment feedback to memory construction recovers part of the drop, and a significant gap remains. If you are measuring your own memory loop, shuffle the task order first.
Catastrophic Remembering
Agentic coding READMEs grow without bound in real repositories, stopping only when the repo retires or someone rewrites the file wholesale. This paper traces the cause to imperfect recall and gives the phenomenon a name that inverts the one continual learning is organized around.
Lost in Compaction
Context compaction is now standard in long-running agent systems, and it silently drops the instructions users most expect to persist. This work names that class, Session Constraints, instructions like "do not delete any emails until I confirm" meant to bind behavior for the rest of a session, and introduces COMPINT to evaluate compactors across multi-turn chat, agentic trajectory, and long-horizon research. Current compactors retain only 17% of injected constraints on average, and most leave the task worse off than running it without compaction at all. Retention swings with the compactor, the prompt, the context length, the phrasing, and where the constraint was injected, which is what makes the loss structural rather than a quirk of one setup. The fix is small and does not touch the compactor or the model: an SC-aware extractor running alongside as a plug-and-play module recovers over 90% retention in all three scenarios.
Cracks in the Foundation
You might assume architectural variations within the dense transformer paradigm barely move accuracy, and in the short-context setting you would be right. This work shows four minor decisions, normalization, GQA, pretraining context length, and sliding window attention, each made by at least one of the Olmo, Llama, and Qwen dense families, have a compoundingly negative effect on long-context extensibility. Any one alone is minor, but combining three or more drops downstream long-context performance by up to 47%, and none of it is detectable from short-context loss or validation sets, which is precisely how these choices survive into shipped models. Applying context extension early in pretraining exposes the problem cheaply. After over 170,000 GPU hours the authors release OlmPool, 26 comparable 7B models with checkpoints before and after extension, including several architectures that beat the Llama 3 architecture on long-context extensibility.
Model or Harness
Agent evaluations mostly report system-level outcomes, so a failed run leaves the repair unassigned. The same visible failure might call for model post-training, harness engineering, environment redesign, or benchmark repair, and outcome labels cannot separate those cases.
Zero-Mem
Production memory stacks spend extra model calls on summarizing interactions, writing records, and reranking retrievals. Each of those calls costs tokens and latency, and the generated summaries quietly discard the evidence you later need. This work asks whether structured memory access requires generation at all.
Rehearse
Autoresearch loops propose changes, run full training jobs, and keep whatever improves the metric. Their efficiency depends on judging, before spending a run, whether a proposed modification is likely to work, and this paper studies how that judgment holds up over a trajectory.
ContinualSkillBench
Skill libraries are shipping in agent harnesses on the assumption that writing skills down compounds, and this benchmark tests that assumption directly. ContinualSkillBench covers five domains, each with 100 interconnected subtasks ordered by increasing difficulty and built with deliberate opportunities for cross-task skill reuse. Sequential execution generally improves performance, though the gains vary substantially across models and domains, and maintaining an explicit skill library performs comparably to plain in-context learning on average. Much of the improvement comes from adapting to prior context and feedback rather than from reusable skill abstraction, though explicit skills still pay off selectively on tasks needing reusable procedures or precise outputs. There is a useful diagnostic buried in the results. Less capable models accumulate larger, more fragmented collections of task-specific skills, which is what failed abstraction looks like from the outside.
ACM
Production agents accumulate context every turn. The usual fix compresses on a token threshold and throws the remainder away, so the trigger fires for reasons unrelated to what the agent is working on. Meta and CMU hand the decision to the agent instead.
Filesystem Memory Audited
Deployed agents increasingly keep long-term memory as a directory tree of markdown files they read, write, and reorganize with ordinary file tools. Research had mostly designed bespoke memory representations instead, leaving the default's two working assumptions untested.
Beyond AdamW
Higher-order optimizers have promised faster convergence than AdamW for a while, with computational cost and numerical stability as the standing objections. This NVIDIA work adapts them for large-scale pretraining, identifying instabilities in SOAP at large batch sizes and eliminating the loss spikes with per-step QR orthogonalization and improved preconditioning, then running a unified study of SOAP, Muon, and AdamW under update-RMS matching for fair learning rate transfer. On multi-billion-parameter models trained over trillions of tokens, SOAP and Muon consistently beat AdamW, and at batch sizes up to 100M tokens for next-token prediction they hold stability and quality while AdamW degrades. A layer-wise distributed optimizer compatible with Megatron-LM balances memory and hides communication without approximating the optimizer math.
Role Drift
End-to-end RL improves the accuracy of a multi-module LLM pipeline without constraining how the modules divide labor internally. Harvard and MIT name the resulting failure mode, Role Drift, where a module preserves or improves end-task performance while abandoning its assigned role through shortcuts that system-level evaluation cannot see. Two instances showed up. A decomposer meant to split a question into sub-questions for a separate solver instead plants the answer inside them, and a reader meant to answer from retrieved passages instead falls back on parametric memory. Hold the decomposer to its role and 86% of the apparent RL gain disappears. Role Anchor, the proposed regularizer, preserves how the role prompt shifts a module's next-token predictions relative to a neutral prompt, and gradient analysis suggests it reduces alignment with the drift direction rather than simply suppressing learning.
The Self-Speculating Agent
Agents spend a large share of wall-clock time waiting on tool results. Speculation hides that latency by predicting and pre-executing the next call, but external draft models and cached traces model a different policy, so they miss too often to help. UC Santa Barbara and LinkedIn identify this speculator-agent gap and unify both roles in one model. It runs in agent mode to solve the task and in speculator mode to predict its next tool call from a partial trajectory, fully reusing the prefix KV cache. Joint agent-speculator reinforcement learning derives speculation targets from the agent's own rollouts and alternates updates between the two modes. Next tool-call Hit@1 rises from 44.1 to 61.2 for Qwen3-4B and from 48.9 to 66.3 for Qwen3.5-4B, with agent task success preserved.
From Memory to Skills
Most agent memory systems retrieve past traces as passive context, so hard-won experience never becomes something the agent can directly execute. MSCE, a training-free memory-skill co-evolution framework, instead governs how experience turns into callable skills for long-horizon LLM agents.
PRO-LONG
Long-horizon tasks force a harness to decide what to save from a long stream of observations and how to load it back into context, and richer summaries usually make the exact detail you need harder to retrieve. PRO-LONG sidesteps this tradeoff with programmatic memory.
Bad Memory in Agents
Persistent memory is what makes an agent useful across sessions, and it is also a place an attacker can leave something behind. This work evaluates prompt injection from memory files in Claude Code and OpenAI Codex, across Claude Haiku 4.5, Claude Opus 4.7, GPT-5.2, and GPT-5.5. The finding is uneven but sobering: it is hard to make an agent overwrite its own memory using untrusted external content, but payloads already planted in those files reliably attack current and future sessions, with attack success and persistence varying widely across systems, models, adversarial goals, and multi-session sequences.
Frontier Models Struggle to Copy
Frontier models can write proofs yet stumble on faithfully copying a long block of text that sits well within their context window. This paper traces the failure to 1D positional encodings, whose inductive bias favors a copying shortcut based on matching local context rather than carefully locating the corresponding input positions. The fix is 2D-RoPE, which lays text out on a 2D grid and gives each token a row and a column ID, so copying becomes retrieving tokens at a fixed column offset. Shallow Transformers with 2D-RoPE copy perfectly at input lengths hundreds of times longer than those seen in training.
RoboTTT
Recent robot foundation models run on single-step or short-history context, a strange way to attempt a five-minute assembly task. RoboTTT, from NVIDIA with Stanford and UT Austin, integrates test-time training into vision-language-action policies to scale visuomotor context to 8K timesteps, three orders of magnitude past prior policies, without growing inference latency. The longer context unlocks one-shot in-context imitation from human video, on-the-fly policy improvement, and robustness to perturbations. It improves overall performance by 87% over a single-step baseline, fully completes a ten-stage assembly task that no baseline finishes, and gains 62% from pretraining with 8K rather than 1K timesteps.
Self-Improving Agents Survey
Self-improving agents are moving from research demos into deployed systems, and this survey gives the trend a clean formalism. It frames a modern agent as a foundation model coupled with an operational scaffold of prompts, memory, tools, and control logic, then treats self-improvement as a self-induced update that commits changes to either the weights or the scaffold.
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.
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.