The author argues that while LLMs write ordinary programs with ease, they often take a roundabout route — building complex symbolic reasoning instead of simply calling another model — when designing systems that use LLMs themselves, and explores possible causes in training data and RL environments.
This Hacker News thread asks developers to place themselves on a scale from autocomplete-only use to fully hands-off agents, and share how far they have actually gone with AI coding.
The new Ollama release improves ChatGPT CLI restore and graceful restart, and refines details such as OpenCode settings and memory mapping on Vulkan integrated GPUs.
Alibaba officially launched its new-generation foundation model Qwen4-Max on November 16, with stronger agent and coding capabilities moving toward real-world, all-scenario applications.
S1-Evo shows that typed System-1 decision models deployed for agent routing, moderation, and authorization can still be bypassed by carefully crafted evidence-competition attacks.
When an agent is about to cancel a booking or delete a file for a user, this paper digs into what actually keeps it from acting on a wrong assumption — the information its tools return, or the user's own words.
This work quantifies uncertainty for multimodal LLMs with an incoherence-adjusted semantic volume, so plausible but wrong outputs are not mistaken for reliable ones.
This work aligns speculative-decoding draft models with parameter- and data-efficient adaptation, easing the performance drop draft models suffer after target models are fine-tuned.
After an API migration, code fixed by an LLM can run without errors yet produce wrong results — and standard release checks miss these silent failures entirely, the research finds.
JudgeMix uses judge-guided mixed-rollout data synthesis to ease the known train-test mismatch in supervised fine-tuning on expert trajectories for LLM agents.
This paper shows that hard-deleting facts a memory system judges stale can actually hurt downstream accuracy, so the real question is when forgetting is safe.
The study places Gemma 3 12B in a simulated courtroom to test whether sustained gain-loss framing can push a long-horizon agent toward market-manipulating collusion.
This study has LLM agents use hill-climbing search to autonomously determine crystal structures from powder X-ray diffraction data, clearing a key bottleneck for self-driving materials labs.
This study disentangles three factors often confused in multi-agent evaluations, clarifying whether gains come from coordination between agents, ensembling results, or simply more test-time compute.
This study evaluates LLMs' risk propensity across contexts and task demands, systematically measuring whether models lean risk-averse or risk-tolerant.
BigCodeBench evaluates LLM code generation with tasks involving diverse function calls and complex instructions, closer to real development than benchmarks that only test simple snippets.
This work proposes a portable screen that first checks whether an LLM's confidence signal actually varies with correctness before using it for abstention, routing, and escalation decisions.
DeltaBox lets LLM-powered agent workloads checkpoint and roll back sandbox state at millisecond scale, supporting high-frequency state exploration like test-time tree search and reinforcement learning.
QVal grades the quality of dense supervision signals on long-horizon agent tasks, since rewards based only on final outcomes can't tell which intermediate steps went well.
REFORGE benchmarks LLM reverse-engineering capability through decompiled binary function naming, responding to real threat-intelligence reports of LLMs being used for reverse engineering.
This work enables cloud and edge LLMs to learn collaboratively by accelerating personalized bilevel optimization, supporting distributed training for physical-world intelligence.
Agentic-OPD recovers agent capabilities forgotten during training by using on-policy distillation, countering the common practice of keeping only the best-validation result in RL for agentic tasks.
Aimed at real deployments, this study tests whether leading hallucination detectors still hold up on small, quantized on-device models, giving practitioners measured evidence for choosing one.