Research

He & Yu · 2026

Synthesized September 18, 2026 · Proves that quorum size is not a safety dial in multi-agent approval systems. When reviewers share upstream data sources, their votes collapse onto a single corrupted root. Three formal results show adding more reviewers does nothing when the new reviewers read the same inputs. Introduces the Structural Epistemic Cut (kappa-E) and a Dependency-Aware Quorum Controller.

Wu & Canedo · 2026

Synthesized September 16, 2026 · Moves cross-episode agent caching from a storage problem to a protocol problem. Servers attach version stamps and cacheability hints; clients evict stale entries deterministically. Row-level invalidation recovers 29-33% of token cost; table-level invalidation destroys recovery entirely. Compliance is a model property: identical signals yield 100% first-try compliance on Haiku 4.5 and under 11% on Sonnet 5.

Lin, Ang, Zhu, Ding & Zhou · 2026

Synthesized September 15, 2026 · Scroll replaces the growing text-buffer context model with a sandboxed Python namespace, an Event Log, and an eviction landmark index. Context management becomes a coding task. 86.7% on LOCA_256K, +37.4pp over the best published long-horizon agent result, with a single backbone. Lin, Ang, Zhu, Ding & Zhou, 2026.

Leins, Pelleriti, Gonnermann-Müller & Pokutta · 2026

Synthesized September 14, 2026 · When every method including the single-call baseline is optimized equally, multi-call orchestration gains 4.6 pp over optimized CoT at 2–4x the token cost. No evidence harder tasks justify the overhead. Strong method-by-backbone interaction means a winning configuration on one model may not transfer.

Jolicoeur-Martineau, Sukthanker, Cameron & Gervais · 2026

Synthesized September 13, 2026 · A missing-baseline study: SWA with attention sinks needs no post-training and matches or beats post-trained linear attention on general tasks, outperforming it 2–10x on long-context reasoning benchmarks.

Xiao, Sun, Wu, Hui et al. · 2026

Synthesized September 12, 2026 · A supervisor-worker harness that aborts doomed runs mid-execution and distills lessons into skills and memory the same run uses immediately. Output tokens fall 42.9–47.4% while successful evaluations per million output tokens rise 110.3–134.0%. First place in 5 of 6 benchmark configurations across two backbone models.

Hong, Chi, Li, Wang, Gao et al. · 2026

Synthesized September 11, 2026 · A three-stage evaluation closes the loophole where coding agents copy the original code forward and pass all tests without migrating. Only 5.4% of 520 frontier-model runs across 8 models complete a whole-repository migration end to end. Among audit-passing runs, 58% reach 99% of fixed tests and only 26% reach 100%.

Anand & Chattaraj · 2026

Synthesized September 10, 2026 · Engineered decoy tools seeded into agent tool sets reveal that canary susceptibility varies 36x across eight models. Capability tier does not predict selection safety. Frontier models have one specific blind spot: capability mirages, tools that look applicable but exceed the agent's actual scope. The framework is released and requires no fine-tuning.

Parupudi · 2026

Synthesized September 9, 2026 · A fragility audit of 12 LLMs across 3,679 benchmark items and 26 harness configurations finds that config-fragile items carry 95.7% of adjacent-model score gaps. Four of 12 models can claim rank one depending on harness choice. Benchmark compression selects for fragile items. Scoring mode, not option ordering, is the load-bearing variable.

Huang, Wang, Han, Yan, Chen et al. · Google · 2026

Synthesized September 6, 2026 · A programmable harness layer wraps an existing agent training environment to target the specific weaknesses of whichever agent is learning in it, without touching the verifier or internal logic. Up to 9.0 points of held-out gain and 9.8% fewer execution steps across five benchmarks in four domains, beating the original environment and domain-specific generation pipelines.

Chi, Li, Hong, Wang et al. · 2026

Synthesized September 5, 2026 · A benchmark for whether LLM agents redesign training algorithms, not just tune parameters. 29 configurations, 10 domains. Best system scores 0.250 (shipping baseline is 0.1). Mechanism-change submissions average 0.226 vs. 0.126 for parameter-only. Raising reasoning budget moves mechanism attempts from 8% to 64%.

Xu, Yan, Chen & Kechadi · 2026

Synthesized September 4, 2026 · A frozen untrained model run through the standard self-improvement pipeline produces apparent capability gains. Seven measurement failures each invert a published finding when a null is present. External distillation improves hard problems; self-training does not pass a corrected null test.

Liu, James, Wang, Xiao & Lin · University of Manchester, University of Sheffield & SUFE · 2026

Synthesized September 3, 2026 · A tool retriever trained on one query source style drops from 86.1% to 22.3% coverage when mixed-source queries arrive in production. TF-IDF fingerprinting detects the style shift; a routing guard with 20 examples per source restores coverage to the single-source baseline without retraining.

Kolukuluru, Ashok, Arora, Dernoncourt, Rossi, Lipka et al. · Adobe Research · 2026

Synthesized September 2, 2026 · The first stage-aware pruning comparison in long-horizon research agents finds pipeline position dominates scoring method. Pre-retrieval pruning cuts tokens up to 73% with little quality loss. No method wins across quality, efficiency, and faithfulness simultaneously.

Jiang, Huang, Xing, Wang, Liu, Li et al. · 2026

Synthesized September 1, 2026 · Skills help agents by stabilizing action sequences in 65.7% of helpful cases, not by injecting missing knowledge (4.5%). Retrieval precision collapses from 29.6% to 3.3% as the skill pool grows from 5 to 100 entries, a bottleneck independent of content quality.

Chen, Liu, Zhu, Dou et al. · 2026

Synthesized August 31, 2026 · REDAgentBench checks service receipts instead of transcripts and finds trajectory-only evaluation systematically undercounts safety violations. In nearly one in five confirmed violations, the agent stated the relevant constraint before ignoring it. Training-free policy reminder at the action boundary cut violations by over 70 percentage points in matched replay.

Li, Yang, Tan, Huang et al. · 2026

Synthesized August 30, 2026 · Seven frontier models on 36 long-horizon R&D tasks, graded by what happened inside each run: solution framing, execution, and feedback control. Agents behave like engineering optimizers. Run-to-run variance is high. Experience transfer is near-zero. Pass rate alone hides all of it.

Arjmandi · 2026

Synthesized August 29, 2026 · Standard long-context benchmark truncation drops content from the middle of prompts, removing task-relevant signal alongside filler. Distractor-aware truncation preserves those spans and keeps accuracy at or above the full-context baseline, with statistically significant gains on BABILong for Claude Haiku 4.5 and Sonnet 4.6. Some published accuracy-vs-context-length curves may partly reflect truncation methodology rather than true window limits.

Cai, Zhao & Li · East China Normal University · 2026

Synthesized August 28, 2026 · Controlled capacity sweeps across a three-role hierarchical search system show decomposition is the bottleneck: scaling the delegation backbone lifts exact-match ~11 points, while scaling execution sub-agents moves it ~2.6. A 1.7B distilled executor matches a frontier sub-agent at 37% fewer sub-agent tokens.

Weinberger & Hozez · 2026

Synthesized August 27, 2026 · A preregistered benchmark isolates prompt wording as a causal driver of reasoning-model cost in coding agents. Open-ended exploration cues inflate deliberation tokens 2.4–7.4x with no correctness gain. A bounded template with explicit scope, acceptance criteria, and a stop condition keeps spend at baseline.

Xiao, Zhu, Zhang et al. · 2026

Synthesized August 25, 2026 · Researchers eliminated LLM calls from all agent memory operations except final question-answering. Raw traces organized as an entity-context graph plus a temporal hierarchy, combined deterministically per query. The result is 57.6% lower memory-operation time cost versus the fastest baseline at matched answer quality, with zero LLM tokens consumed on memory operations themselves.

Feng, Zhang, Cheng & Qi · 2026

Synthesized August 24, 2026 · Researchers reframe tool acquisition in LLM agents as a stopping problem sitting on top of the retriever's ranked output, and prove that score-only cutoffs (fixed-k, similarity threshold, score-per-cost ratio) are suboptimal once tool costs are heterogeneous. CAM-DF, a training-free stopping rule trained on the offline gap between stopping now and the best continuation, cuts tools per query by 37% at matched task success across 1,343 tasks in five tool-use domains and beats predict-then-threshold in 20 of 20 permutations at cost heterogeneity d = 1.0.

Orgad, Barez, Haklay et al. · 2026

Synthesized August 23, 2026 · A position paper from a coalition of NLP researchers argues the field should grade interpretability findings by whether they enable a concrete, testable intervention, not by elegance of internal explanation. Proposes a two-axis rubric (concreteness x validation) and identifies five domains where actionable interpretability outperforms behavioral evaluation alone.

Akhauri & Abdelfattah · Cornell University · 2026

Synthesized August 22, 2026 · Cornell researchers trained a small policy network to select attention sparsity, MLP pruning, and quantization bit-width independently at each decode step, using a frozen base LLM's hidden state as the signal. Up to 7.3-point MMLU gain over uniform-budget allocation at matched FLOPs, with a better quality-efficiency Pareto front than any static efficiency setting produces.

Gollapudi, Gupta, Singhal, Min et al. · 2026

Synthesized August 20, 2026 · Researchers named attention dilution as the specific failure mechanism behind long-context retrieval collapse: irrelevant documents inflate the softmax denominator and drain normalized attention mass from the gold document even when its raw score stays high. Two length-aware fixes restore accuracy; BlockSearch (0.6B parameters) outperforms a competing retriever seven times its size on MS MARCO and NQ, and length-generalizes up to 10x beyond training.

Zhou, Zhao et al. · Baidu · 2026

Synthesized August 16, 2026 · Baidu researchers attached human labor costs to every office benchmark task and found every model tested was faster and cheaper than a human worker, yet none produced deliverable-quality output. Quality-per-dollar is the metric that makes the gap visible.

Weinberger & Hozez · PointFive · 2026

Synthesized August 12, 2026 · PointFive researchers ran 2,848 metered Claude Code sessions and found that removing 38.4% of tool-output tokens raised billed costs by 6.8%. Prompt-cache traffic accounts for roughly 87% of the four-component cost of an agent session, and compression feedback loops can transfer cost rather than eliminate it. Token count is an unreliable proxy for the provider invoice once caching is in the picture.

Bousetouane · ProofAgent.ai / University of Chicago · 2026

Synthesized August 11, 2026 · A seven-criterion scoring harness for agent context quality, isolated from behavioral metrics before any tests run. In 300 controlled sessions holding the model fixed, critical failures fell from 4.11 to 1.33 per session when context moved from poor to structured: a 68% reduction before any safety hardening was applied. Each criterion predicts its matched failure class.

Li, Hou, Jing, Tang & Dong · Tsinghua University · 2026

Synthesized August 10, 2026 · Tsinghua researchers trained context compaction into the RL policy for long-horizon agents. Cross-trajectory credit assignment and token-level loss normalization let agents learn what to preserve when the context window fills. SWE-bench Verified 66.8% (+7.0pp) and Terminal-Bench 2.0 24.5% (+3.1pp) on GLM-4.5-Air under a fixed context budget.

Zhou, Leonardis & Feng · University of Birmingham · 2026

Synthesized August 9, 2026 · Static agent routing assigns steps by function labels, ignoring per-step competence differences. Agora replaces routing with an incentive-compatible auction: candidate models bid for each reasoning step based on rectified self-confidence, and honest bidding is the dominant strategy. Outperforms single-model, routing, and cascade baselines across five benchmarks; MMLU-Pro 71.9% vs 68.1% best single-backend.

Xiang, Chen, Tang, Wei, Ning, Lin, Zhang & Su · 2026

Synthesized August 8, 2026 · A five-task benchmark separating retrieval quality from judgment quality in memory-equipped agents reveals that existing memory systems often increase sycophancy relative to no-memory baselines. Most failures trace to memory use rather than retrieval error: the agent retrieved the right memory, then deferred to it when it should have pushed back.

Yin & Feng · University of Tennessee, Knoxville · 2026

Synthesized August 7, 2026 · Agents run 5–10x more compute than simple tasks need, wasting 80–90% of their budget on cognitive redundancy. E3 (Estimate, Execute, Expand) classifies difficulty before starting, caps file reads for simple tasks, and expands scope only if verification fails. On MSE-Bench: 100% success matched at 85% lower cost, 91% fewer tokens, 92% fewer files read.

Li, Liu, Zhang et al. · Beijing Jiaotong University & Tencent Weixin AI · 2026

Synthesized August 6, 2026 · 67 multilingual long-horizon workplace tasks across five domains show SOTA agents degrade sharply when language switches across workflow steps. A linguistic error at step 2 propagates into every downstream output, and the harness accounts for 0.08–0.21 Pass@1 variance. The LLM semantic judge correlates with structural correctness at r=0.18 (r=-0.04 at high confidence).

Karamchandani, Nagasubramaniam, Zhu & Wu · Pennsylvania State University · 2026

Synthesized August 5, 2026 · FARMA shifts memory poisoning from facts to reasoning traces, using evasive phrasing to defeat keyword filters and self-referential amplification to defeat consensus defenses, reaching 100% attack success rate. The paired SENTINEL defense uses structural trace analysis to drop that rate to 0% with zero false positives across 326 benign traces.

Virk, Edds, Xia & Zhang · University of Illinois Urbana-Champaign · 2026

Synthesized August 4, 2026 · A Shepherd Agent with global context steers isolated Search Agents in separate git branches, producing better or comparable solutions to SOTA LLM-guided evolution on 13 of 15 open-ended optimization tasks by keeping the swarm exploring distinct high-level approaches.

Xu, Sun et al. · Alibaba & ShanghaiTech University · 2026

Synthesized August 3, 2026 · NapMem organizes user history into a four-level memory pyramid and exposes each layer as a callable tool. An RL-trained agent navigates between granularities based on intermediate evidence, producing better answers with fewer unnecessary tool calls than passive retrieval baselines.

Chen, Duan, Sun et al. · HKU MMLab · 2026

Synthesized August 2, 2026 · A 400-task benchmark running agents in live Docker containers that grades step-by-step checkpoints across five capability dimensions, revealing that framework choice moves agent scores as much as model choice does.

Biswa Sengupta · 2026

Synthesized August 1, 2026 · A frozen-base agent accumulates steering adapter updates from five verifier loops, then gates every self-modification behind an anytime-valid statistical test before it lands. +4 and +5 resolved instances on a 52-instance SWE-bench Verified subset across GLM 5.2 and GPT.

Wang et al. · Notre Dame, Stanford, Yale et al. · 2026

Synthesized July 31, 2026 · Step-wise reasoning collapses into a greedy policy on long tasks; FLARE's bounded lookahead raises recovery rates from 5.4% to 29.7% and lets LLaMA-8B outperform GPT-4o on planning benchmarks.

Ding, Xie, Wei, Li & Ding · Renmin University of China & Alibaba Group · 2026

Synthesized July 30, 2026 · A tool-using LLM agent mines its own successful trajectories, promotes recurring multi-step sequences into callable higher-order SOP tools, then merges and prunes them through a three-phase lifecycle parallel to ML training. Task success improved 2.5–13.4 percentage points on ACEBench depending on backbone, with fewer interaction rounds per task.

Reddy, Challaram & Basu · 2026

Synthesized July 28, 2026 · On tau-bench airline, 78% of a budget agent's failures were silent wrong-state writes: tasks the agent reported done, tools accepted, but resulting state violated business policy. Four deterministic read-only pre-execution gates lifted success from 29.6% to 42.0% on gpt-4o-mini (+12.4pp, P=0.0012) and 61.2% to 71.6% on gpt-5.2, with negative controls confirming gates help only where tools are policy-permissive.

Lv, Wu, Zhu, Cheng & Guo · Nanjing University LAMDA · 2026

Synthesized July 27, 2026 · OpenAgent formalizes four shift types (query, action, observation, domain) that break trained tool-using agents, diagnosed against a four-tier failure hierarchy. Both SFT- and RL-trained agents degrade under open-environment shifts; Perturbation-Augmented Fine-Tuning improves robustness. ICML 2026.

Lin, Wang et al. · Northwestern MLL Lab · 2026

Synthesized July 26, 2026 · Frontier agents waste 28–64% of tokens on tasks they ultimately fail, and task skill barely predicts budget-awareness (r = 0.35). BAGEN frames budget prediction as a metacognitive skill and trains early-stop and alert behaviors via SFT+RL, reaching 47% interval coverage.

Marozzo & Liò · University of Calabria & University of Cambridge · 2026

Synthesized July 25, 2026 · An evidence-first investigator agent scores description ambiguity, generates competing hypotheses, asks targeted clarifying questions, and commits only when one explanation dominates -- beating direct-prompting and reasoning-only baselines on a benchmark of solved forum threads

Wang, Wen et al. · Tencent · 2026

Synthesized July 24, 2026 · Tencent researchers prove mathematically that bi-encoder embeddings cannot learn compatibility signals for agent skill sets, then build a two-stage retriever mining LLM rejection decisions. Set-level compatibility accuracy jumps 13 percentage points on a new 5,696-query benchmark

Zhang, Xu, Li, Zhang, Jiang, Zhang & McAuley · UC San Diego · 2026

Synthesized July 14, 2026 · A systematic sweep over 4B-to-284B models and three retrievers maps when observation masking helps long-horizon search agents and when it collapses performance: an asymmetric inverted-U driven by model-retriever interaction, not either factor alone

Xiao, Ma, Chen, Chen, Atreya, Chen & Ordonez · Rice University / Meta · 2026

Synthesized July 10, 2026 · RA-RFT trains a retriever by expected reasoning utility rather than semantic similarity, then fine-tunes via RLVR on the analogous demonstrations. Up to +7.1 points on AIME 2025 over GRPO

Zhu, Li, Lyu, Luo, Yang, Liu, Hui, Yuan, Sun, Su & Shao · Beijing University of Posts and Telecommunications, Shanghai AI Lab & collaborators · 2026

Synthesized July 9, 2026 · Seven agent benchmarks re-run through one standardized harness and fixed ReAct loop across 15 models, 400K rollouts, and 5B tokens. Scaffold choice and environment volatility alone are sufficient to reorder leaderboards

Wang, Qian, Zhang et al. · Boston University, Northeastern University & University of Texas at Dallas · 2026

Synthesized July 8, 2026 · Four frontier LLM agents tackle 51 research artifact deployment tasks in clean environments. Best pass rate: 51.0%. The dominant failure: agents stopping before verifying the artifact actually ran

Liu, Xiong, Zhang & Tang · Emory University & Shanghai Jiao Tong University · 2026

Synthesized July 7, 2026 · A diffusion world model trained on historical interaction logs simulates environment responses for any candidate LLM agent, enabling reliable off-policy performance estimates without live deployment. Solves a structural problem that broke all prior OPE methods for discrete autoregressive agents

Sun, Wang, Song, He, W. Zhang, Y. Liu, Y. Yang & Y. Cheng · Tongji University, Shanghai AI Lab & collaborators · 2026

Synthesized July 6, 2026 · A diagnostic benchmark with 4,209 cases across 22 domains separates planning failure from execution failure and adds infeasibility detection. Across 12 MLLMs, systematic weaknesses in long-horizon planning, tool-noise robustness, and refusal calibration

Jiang, Huang, Cheng, Deng, Sun & Mai · University of Edinburgh · 2026

Synthesized July 5, 2026 · A context index intercepts overlapping prompt blocks across requests, reorders them into cache-reusable prefixes, and deduplicates repeated content with quality-preserving annotations. Up to 3x faster prefill, 36% fewer tokens, zero model changes

Abdelnabi & Bagdasarian · University of Massachusetts CICS · 2026

Synthesized June 24, 2026 · An impossibility result shows no defensive norm can block all prompt injections without also blocking flows the legitimate task needs. Data-instruction separation is a partial mitigation, not a fix

Fadnavis, Kanakaraj & Wyss · Bioscope AI · 2026

Synthesized June 23, 2026 · In mixture-of-agents systems, reading complete reasoning chains from all agents recovers correct answers that majority voting discards, even when every agent agreed on the wrong answer

Xiaohongshu Inc. · 2026

Synthesized June 22, 2026 · A 200-task bilingual benchmark for proactive, multi-turn search agents. The best frontier model scores 30.3 Triplet F1, revealing how far current search agents fall short when handling real-world intent ambiguity

Lin, Li, Song, Jiang & Zhang · ByteDance · 2026

Synthesized June 21, 2026 · A five-component lifecycle for LLM agent skill libraries: create skills from successful tasks, gate admission with unit tests, refine from new examples, retire underperformers. Auto-generated skills hit 68.40% overall and beat the human-skill ceiling on tasks where generation succeeded

Orogat & Mansour · Concordia University · 2026

Synthesized June 20, 2026 · Long-term agent memory is a data-management workload, not just a retrieval problem. Existing vector stores support ingest and retrieve but lack a revise operator, and that gap explains most agent memory failures. GEM formalizes four state-level operators replacing CRUD

Akhauri & Abdelfattah · Cornell University · 2026

Synthesized June 19, 2026 · A lightweight policy reads the LLM's own hidden state at each decode step and picks how sparse, pruned, and quantized that step should be. Up to 7.3-point MMLU gain over uniform compression at matched FLOPs

Orgad, Barez, Haklay et al. · multiple institutions · 2026

Synthesized June 15, 2026 · A position paper proposing two criteria for evaluating interpretability work: concreteness of the intervention enabled and validation that the intervention works. Most current work falls short on at least one dimension

Al-Tawaha, Gu, Niu, Jia & Jin · Virginia Tech, UC Berkeley & UIUC · 2026

Synthesized June 14, 2026 · Memory-induced safety violation rates climbed monotonically with exposure length across 8 memory architectures and 3 deployment scenarios. The same agent, same probes, more memory produced more safety violations. Risk is detectable at retrieval time before generation

Lei, Zhang, Li, Wang et al. · Nanyang Technological University · 2026

Synthesized June 13, 2026 · A compact 8×8 online state matrix uses delta-rule learning to compress past context and inject low-rank corrections into a frozen LLM's attention at each decode step. 1.31× the backbone on MemoryAgentBench, 1.20× on LoCoMo, no base model retraining required

Yang · Independent · 2026

Synthesized June 8, 2026 · An 8-component runtime pipeline intercepts every agent tool call before execution: deobfuscates shell payloads through nine normalization strategies, tracks multi-step attack chains across sessions via RiskChain, and proposes safer alternatives via 37 SafeFix rules. 95.0% verdict accuracy at low-millisecond latency

Marin-Llobet & Ferrando · Harvard University · 2026

Synthesized June 7, 2026 · Two coupled agent loops automate mechanistic interpretability end-to-end: a discovery agent navigates activation space via k-NN graphs to surface interpretable features, and an explanation agent iteratively probes and refines hypotheses to beat one-shot auto-interp baselines

Zheng, Liu, Huang et al. · UMD, UVA & UNC · 2026

Synthesized June 6, 2026 · A coding agent discovers test-time scaling controllers by searching a frozen reasoning trace cache. The found CMC controller cuts tokens 69.5% vs SC@64 at matched accuracy, for $39.90 discovery cost

Xu, Li, Zhao, Wu, Li & Yan · Xi'an Jiaotong University · 2026

Synthesized June 2, 2026 · DDC combines confidence-weighted Bayesian voting and trend-aware path pruning to make self-consistency sampling adaptive. Over 10x token reduction at matched or improved accuracy across five reasoning benchmarks

Sunkaraneni, Beneventano, Neumarker, Poggio & Galanti · MIT & Texas A&M · 2026

Synthesized May 30, 2026 · A committee of nano-scale models hit 76.4% on SWE-bench Verified, matching Gemini 3 Pro and Claude Opus 4.5 Thinking standalone. The framework explains when and why committees of weak models match frontier accuracy: the task needs a local verifier

Sen, Kasturi, Lumer, Gulati, Subbiah et al. · 2026

Synthesized May 28, 2026 · Testing grep vs vector retrieval across four agent harnesses (Chronos, Claude Code, Codex, Gemini CLI) reveals the harness layer shifts accuracy more than the retrieval method; Claude Code favors grep, Gemini CLI favors vector

Wang, Qiu et al. · Princeton, Michigan, Tsinghua et al. · 2026

Synthesized May 23, 2026 · A training-free cold-start router builds experience memory from a small seed set to route queries between plain LLM inference and full agent execution; 60.6% inference time reduction vs always-agent, 28.6% accuracy gain vs always-LLM

Li, Wang, Liu et al. · 2026

Synthesized May 18, 2026 · A training-free two-phase method explores in latent space before switching to explicit chain-of-thought. On Qwen3-14B: 32% token reduction, AIME 2025 accuracy improves from 70.0% to 73.3%

Li, Yang, Wang et al. · 2026

Synthesized May 17, 2026 · A 150+ tool MCP benchmark exposes three reproducible failure modes in frontier LLM agents: tool-retrieval saturation, over-confidence skipping verification, and strategic defeatism

Chao, Bai et al. · 2026

Synthesized May 15, 2026 · A 1,200-query benchmark probes whether frontier LLMs can detect when stored memories have been silently invalidated. Best model scores 55.2%. Implicit conflict is the dominant failure mode

Zheng, von Glehn, Zwols et al. · Google DeepMind · 2026

Synthesized May 14, 2026 · The same base model scores 19% alone and 48% inside a multi-agent workbench with parallel workstreams, stored failure records, and enforced review cycles

Lee, Nair, Zhang, Lee, Khattab & Finn · Stanford University & MIT · 2026

Synthesized May 13, 2026 · Fixing the model and varying only the surrounding harness code produces a 6x performance gap. Meta-Harness automatically searches for better harness code using full diagnostic history, beating SOTA on text classification, IMO-level math reasoning, and agentic coding

Popescu, Gros, Botocan, Pandita, Devanbu & Izadi · TU Delft & UC Davis · 2026

Synthesized May 12, 2026 · 110,000 open-source pull requests from five coding agents reveal that agent code gets merged but churns faster than human-authored code over time

Taneja, Singh & Goel · Georgia Institute of Technology · 2026

Synthesized May 11, 2026 · A 124-person randomized controlled trial found multimodal conversational AI produces better biology learning outcomes than text-only chat or semantic search, explained through two distinct cognitive load mechanisms

Dombrowski, Zhang, Cechnicka, Reynaud & Kainz · FAU Erlangen-Nürnberg & Imperial College London · 2025

Synthesized May 10, 2026 · No state-of-the-art image generator covers more than 77% of its training distribution, and standard metrics like FID can't detect the gap. A new metric (IRS) and a model variant (DiADM) address both problems

Wei, Li & Jiang · Shanghai Institute of AI for Education · 2026

Synthesized May 8, 2026 · A four-stage reasoning workspace that separates cognitive diagnosis, stability validation, affect prediction, and strategy selection before any tutoring response is generated

Tran & Kiela · Stanford University · 2026

Synthesized May 7, 2026 · When you control for thinking-token compute, the multi-agent advantage on multi-hop reasoning largely disappears across five architectures and three model families

Johnson, Ahmed, Lang, Thethi, Zheng & de Souza Santos · University of Calgary · 2026

Synthesized April 19, 2026 · Student developers embedded LLMs as architectural components in two games. The finding: model errors turned into fairness violations, not cosmetic bugs

Shuai Guo · Uppsala University · 2025

Synthesized April 9, 2026 · A mobile strategy game deliberately gives players unreliable AI guidance, reshaping agency from action to interpretation

Pimenova, Fakhoury, Bird, Storey & Endres · U Michigan / Microsoft Research · 2025

Synthesized April 4, 2026 · The first qualitative study of vibe coding reveals a new programming paradigm built on flow and calibrated AI trust

Solyst, Nakigozi, Fong & Shapiro · University of Washington · 2025

Synthesized April 3, 2026 · Two multiplayer games use text-to-image AI to teach teens about bias through play

Li et al. · UBC / Vector Institute · 2026

Synthesized March 25, 2026 · Budget-Aware Value Trees cut AI agent costs by 75% with equal or better accuracy