Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
An October 8 analysis by Engram tested 6 memory systems on 1,536 LoCoMo questions, evaluating retrieval and answer accuracy separately, and found that retrieval success does not guarantee answer correctness. The Agent Me…
OpenCode 2.2.5, the MIT-licensed AI coding agent from Anomaly, shipped on October 30, 2026, and a new guide documents installing it via curl script, npm (@opencode/cli) or Homebrew, connecting models from more than 75 pr…
XAI will begin embedding an invisible statistical watermark in every text response from seven Grok API models — grok-4.7, grok-4.6, grok-4.5, grok-4.3 and all three grok-4.20 variants (reasoning, non-reasoning, multi-age…
Axion 1.0, a fully open agent foundation model, was released with a training recipe of only three stages: two lightweight supervised fine-tuning stages followed by reinforcement learning, with no dedicated midtraining st…
A Qwen3-8B-based long-context agent named MEMENTO, trained to manage its own context through compaction, raised book-level question-answering accuracy from 36.9% to 67.8% while never holding more than 20K tokens in conte…
Identical Gemini 3.6 Flash buyer and seller agents reached agreement in 77.3% of Korean-language negotiations versus 52.5% in English — a 24.8-point gap, Holm-adjusted p = .0039 — across 432 attempted negotiations of 12 …
A new arXiv paper (2611.00450v1) introduces ResMem, a parametric-memory method that cuts the storage retained for teacher construction by 94.7% versus an evaluated kNN pipeline on a Wikipedia subset. Across five question…
A framework that trains a small large language model to predict future reader responses as low-dimensional trajectories, then uses a larger off-the-shelf writer LLM to generate story segments following those trajectories…
All 11 commercial AI-text humanizers on a public leaderboard changed or dropped at least one checkable unit — numbers, dates, person and organization names, or quoted spans — in 35-65% of the October 2026 documents that …
A new arXiv paper (2611.00420v1) introduces Circuit Oracle, an off-the-shelf LLM with a multi-agent harness and task-specific authored skills that answers natural-language questions about a target model by reading its at…
A formal grammar-lens benchmark derived from computation theory found that GPT-5.5 and its Pro variants accepted progressively more novel test data as test-time reasoning effort rose, with acceptance rates of 16%, 40%, 4…
A new arXiv paper (2611.00492v1) comparing human comedy improvisation with multi-agent LLM-generated dialogue finds that large language models can produce coherent, creative scenes but often fail to reproduce key charact…
A new arXiv paper (2611.00536v1) introduces Dual-Axis Policy Entailment (DAPE), a benchmark of 1,000 expert-annotated ordered pairs of Chinese government policy directives that tests whether two directives govern the sam…
A preregistered study of 23 confirmatory tests found that eliciting five weighted estimates within a single response did not uniformly debias language model judgments across anchoring, hindsight, base-rate, conjunction, …
A study on DisasterBench found that majority voting leaves an already-discovered correct workflow unselected on 10-19% of tasks per candidate pool, while structural validation never removed the last correct workflow in t…
A multimodal LLM pipeline named ElectroChemLLM achieved an F-1 score of 0.97 on the battery-electrolyte subset of a manually annotated benchmark of 232 papers, according to the arXiv paper 2611.00638v1. The pipeline comb…
A new arXiv paper (2611.00616v1) proposes Multi-Level Reasoning (MLR), a method that splits long chain-of-thought generation into a high-level planner that emits structured step descriptors and a low-level executor that …
CompSkill, a skill-retrieval method detailed in arXiv paper 2611.01115v1, raises SkillRet NDCG@10 from 64.0 to 70.9 at one eighth of the key dimension by indexing skills with learned 512-d keys and compiling matched same…
A study posted as arXiv:2611.00672v1 found that flipping the recorded decisions of near-copy past cases from DENY to APPROVE raised approval rates by roughly 30 to 55 percentage points in DeepSeek V4.1 Flash, GPT-5.6 Lun…
A new arXiv paper (2611.00726v1) constructs a budgeted evasion attack against KGW text watermarks by replacing token-event evidence through ordinary phrase rewrites, using an independent model's proposed alternatives and…