Large Language Models
1
Muir: Linear Attention with Sensitivity-Allocated Gating and Matched-Filter Readout
Muir is a new linear-attention architecture that allocates gating by sensitivity and reads its recurrent state with a matched-filter readout. From 166M to 1.34B parameters it lowers perplexity by 1.8–2.0% vs. GDN-2 and 2.5–3.6% vs. a Transformer, with up to 1.9 points higher commonsense accuracy. It extrapolates 2,560x from 4K training length to 10M-token streams without perplexity drift and reaches near-100% passkey retrieval up to 1M tokens, while matching GDN-2 throughput and training 2.7x faster than a Transformer at 32K context.
Why it mattersLinear-attention models usually trade recall for efficiency; Muir claims to beat both a strong gated-delta baseline and a Transformer on recall while extrapolating to 10M tokens. If the gains hold at larger scale, it strengthens the case for recurrent layers in long-context and agent workloads.
arXiv ↗PDF ↓
2
The Gating Mechanism of KV Cache Eviction
The paper explains why learned gates for KV cache eviction tend to fail: a gate inside the softmax normalizer is the right design, but trained on the language-modeling loss it collapses to a constant, a no-op. The fix is to keep dense attention out of the gate's reach and train against the dense model's predictions, recovering evicted entries with parameter-free centroids inside the softmax. At 75% KV compression on Qwen3-8B it scores 90.14 on RULER 8K and 46.03 on LongBench, only 3.72 and 1.36 points below the full cache.
Why it mattersKV cache memory is a key bottleneck for long-context serving and agents. Beyond a strong 75%-compression result, the paper gives a clear diagnosis of why naive trained eviction gates collapse, which is useful guidance for anyone building learned cache policies.
arXiv ↗PDF ↓Code ↗
3
When Bigger Becomes Brittle: Why Compression Hurts More as Language Models Scale
The study examines how quantization and pruning affect in-context retrieval, which is carried by a small set of attention heads. It splits the loss from ablating these heads into what the rest of the network restores, what the responsible components account for, and what reaches the output. For comparable damage, small models restore most of the loss while large models keep most of it, because compensation weakens with scale; compressing the heads directly shows the same pattern. Compression that is nearly harmless in small models can therefore be costly in large ones.
Why it mattersCompression recipes are often validated on small models and then applied to large ones. This work argues that this extrapolation is unsafe for in-context retrieval, so compressed large models need their own retrieval-focused evaluation.
arXiv ↗PDF ↓
4
Synthesizing symbolic models of human and machine reasoning at scale
The authors introduce Behavior-Conditioned Symbolic Synthesis (BCSS), which automatically searches for interpretable, deterministic symbolic programs that reproduce observed behavior, and apply it to humans and machines on hundreds of reasoning problems. Both tend to oversimplify, adopt rules that only partly explain their observations and misinfer dependencies between features. Machines' reasoning effort predicts how much they simplify, suggesting both people and models satisfice under resource limits.
Why it mattersIt offers a scalable, interpretable way to describe how reasoning models actually solve problems, rather than only scoring their answers. The shared human-machine biases it finds are directly relevant to how reasoning budgets are set.
arXiv ↗PDF ↓
AI Agents
5
Zero-Shot Coordination among LLM Agents
The paper brings zero-shot coordination, coordinating with partners never seen before, from multi-agent RL to LLM agents. Even in minimal, controlled settings, frontier LLM agents struggle to coordinate, mainly because they misunderstand the coordination problem and reason poorly about their partner's beliefs. LLMs add semantic information as an axis for coordination but fail to exploit it; the authors propose coordination-friendly definitions (CFDs), which can be discovered automatically, to make coordination more robust. Code is released.
Why it mattersAs agents from different vendors increasingly interact, coordinating with unknown partners becomes a practical requirement, not a toy problem. This paper gives a clean testbed showing current frontier models are not there yet, plus a concrete mitigation.
arXiv ↗PDF ↓Code ↗
6
EnvOpt: Automated Optimization for Agent Environments
Rather than optimizing the model or the prompt, EnvOpt automatically optimizes the environment the agent works in. For Claude Haiku 4.5 on AutomationBench it cuts cost by 72.7% and raises pass rate by 14.34%; the same optimized environment cuts Opus 5's cost by 68.1% with a 1.34% gain. On top of Meta-Harness it cuts cost a further 83.4% and adds 11.17% accuracy, and across 5 benchmarks cost falls while pass rate rises or holds.
Why it mattersIt shows that the environment (tool interfaces, observations, structure) is a large, model-agnostic lever on agent cost and accuracy. Because the gains transfer across models, environment optimization could become a standard step next to harness and prompt tuning.
arXiv ↗PDF ↓
7
MCPHallu: Benchmarking Reasoning, Execution, and Memory Hallucinations in MCP Agents
MCPHallu has 358 tasks across five capability domains on 36 production MCP servers in isolated containers, each targeting one of four failure types: Branch Collapse, Unreachable Goal, Tool Misuse or Context Forgetting. Fourteen frontier agents from eight vendors, over 5,012 trajectories, average only 0.642 out of 1. Failures concentrate on unreachable goals (average 0.458, 23.2% pass), where agents keep calling tools or claim success; rankings shift by 6.9 of 13 positions on average across failure types.
Why it mattersMCP is now the default way agents reach tools, and this benchmark tests hallucination on real servers rather than mocks. The finding that agents rarely admit a goal is infeasible is a concrete reliability risk for production deployments.
arXiv ↗PDF ↓Code ↗
8
Who Should Know What? Empirical Information Design for Multi-Agent LLM Economies
An empirical study of how much information to disclose to LLM agents that trade, bargain and bid against each other. Effects are highly model-dependent: revealing exact valuations lowers GPT-6 Sol's realized gains (98% to 91% of maximum) but raises Claude Sonnet 5's (72% to 86%). In bargaining, exact information makes Grok 4.7's deals less balanced while Claude Opus 5.5 splits gains evenly; in double auctions, full disclosure inflates demands and breaks markets, cutting welfare for Gemini 3.1 Pro (97% to 78%), Claude Opus 5.5 (94% to 64%) and Grok 4.7 (97% to 62%).
Why it mattersAgent-to-agent markets are moving from thought experiment to product. The results show there is no universally good disclosure policy: model choice and information design have to be optimized together.
arXiv ↗PDF ↓
9
Terminal-style Mathematical Modeling Bench: The First Agentic and Verifiable Benchmark for Mathematical Modeling
TB-MathModel has 18 real contest-style modeling problems from CUMCM and MCM (2023–2025); each runs in a containerized terminal and the agent returns one JSON artifact. Feasibility is checked by a programmatic gate over six invariant forms plus 624 Lean 4 obligations, with infeasible answers scored zero, and each task is scored against a panel of human outstanding papers. The best of seven multimodal entries reaches 64.58% O-Eval, while a $2.54 open model reaches 59.81% at about 1/100 of the cost.
Why it mattersOpen-ended modeling tasks are hard to grade, so most agent benchmarks avoid them. Combining terminal execution with Lean-checked feasibility gives a verifiable way to evaluate this kind of applied reasoning.
arXiv ↗PDF ↓Code ↗
10
Autonomous AI Agents Systematically Diverge From Human Preferences in Proof Formalization
The study compares where AI agents and humans choose to work when formalizing proofs in Lean, using nine major human collaborations with 4.7M definitions over more than 6K commits. Compared with humans, agents are far more drawn to explicitly marked gaps and relatively avoid densely connected proof centers and complex or surprising goals. They are relatively attracted to surprising proof bodies, which they tend to regularize toward baseline.
Why it mattersAs agents take on more of large formalization projects, where they choose to work shapes the result as much as whether they succeed. The systematic gap with human priorities matters for how such collaborations are steered.
arXiv ↗PDF ↓
