Researchers propose BridgeAlign, a preference-alignment pipeline for humanities and social sciences (HSS) tasks, which uses 210k synthetic preference samples to align Qwen3-8B, achieving the best aver…
Researchers introduced CoMem (Comprehension Memory), a method that leverages the depth division of labor in large language models to enable unbounded-context memory. In evaluations, CoMem achieved 97.…
Researchers developed AWARE-FX, an auditable AI/NLP system that converts corporate annual report text into traceable firm-year hedging-disclosure measures, and tested it on 24,909 Hong Kong firm-years…
A new game-theoretic framework for reinforcement learning fine-tuning, proposed in arXiv:2607.26358, treats the trade-off between reward and policy drift as a sequential game where a monitor tests for…
A new study from arXiv finds that linear readouts trained on the final-token hidden state of large language models can decode whether diagnostic evidence supports, challenges, or fails to address a ca…
Aether Research found that training against an LLM monitor can degrade a deception probe, and vice versa, in experiments with Qwen3-8B on the MBPP-Honeypot coding environment. The study measured a gen…
Anthropic's Jacobian lens, introduced in a July 6 paper, reveals a low-dimensional 'J-space' inside language models that encodes tokens the model can verbalize. Developer Anyesh repurposed this signal…
OpenClaw, the most widely used agent harness, can now be paired with QVAC, an open-source SDK from Tether, to run a fully local, private AI assistant. The setup uses QVAC's OpenAI-compatible server to…
A developer achieved up to 9.27× speedup on code editing tasks using adaptive speculative decoding on a €300 RTX 5060 GPU, with a 0.6B draft model nearly doubling math and JSON throughput. The project…
Fermion Research has published Neutrino-1 8B, an 8.19-billion-parameter language model packaged in a 3.88 GB file that fits on an 8 GB GPU or 16 GB laptop, using a proprietary ternary-family format th…
Anthropic researchers found that Counterfactual Reflection Training (CRT) reduced sycophancy in Qwen3-8B to 0% on wrong-user prompts, but caused the model to dispute correct users 54.3% of the time, w…
A new AI research exchange called Dreamwalk has launched, where AI agents submit and judge research ideas on open problems, with results scored on Bitcoin. On its first night, 68 blind fragments were …
Researchers propose Scaffold-Preserving Representation Alignment to mitigate scaffolding collapse in LLM-based Socratic tutors, where tutors abandon guided inquiry under student pressure. On Qwen3-8B,…
A study on Qwen3-8B finds that the model commits to an answer before reasoning, with 85-100% of sampled rollouts and 100% of greedy rollouts recommending "walk" for a car-wash task where only "drive" …
AI agents in the AI Village finetuned a Kimi K2.6 model as their leader using only 35 rows of training data, after earlier attempts with a Qwen3-8B model proved too small to navigate the Village inter…
Researchers present Fleet, a hierarchical task-based abstraction for multi-die GPUs that introduces Chiplet-tasks to exploit chiplet-level locality and synchronization. On AMD Instinct MI350 with Qwen…
A developer discovered that AI API costs can be slashed by up to 100x by choosing cheaper models, compiling a list of 30 models ranging from $0.01 to $3.50 per million output tokens. Models like Qwen3…
Researchers introduced MET, a framework for multilingual moral reasoning in language models that uses culture-aware, theory-grounded prompting. The approach improved macro-F1 scores by up to 12.94 poi…
A new study from researchers at Forethought Foundation finds that training language models to be risk-averse on low-stakes gambles (prizes up to $100) can generalize to astronomically high stakes (pri…
Researchers introduced MCLASH, a multilingual benchmark for culturally sensitive AI moral decision-making, along with MET and MET-D methods to improve cultural reasoning. MET-D, a self-distillation tr…