Inkling 2 Small produces 316 profanity instances in benchmark runs
Researchers at Vals AI found profanity in 50 of 2,655 model runs across 13 benchmarks, with 241 instances from two spiraling runs.
Stories per day · 30 days
7 RSS
Relatedopenai3multimodal2reasoning2reinforcement-learning2tool-use2actor-critic1adversarial-reasoning1agent-cost1agentic-systems1ai-benchmarks1alignment1api1apple-m41astra1autoregressive-models1causalcot1compaction1contrastive-decoding1
55 stories, newest first · one card per story
Stories tagged #chain-of-thought
CoT reasoning generalizes via shared predictors in LLMs
CoT generalizes OOD by composing reusable ID transitions with risk bounded by O(k) coefficient.
BrainwashedLLM automates RL-based Chain-of-Thought jailbreaks for LRMs
BrainwashedLLM automates discovery of reasoning-based jailbreaks that transfer zero-shot to larger LRMs.
Compacting reasoning history shortens future generation and can cut agent cost
Open-weight models rewrite earlier thoughts into compact state records before tool results, without retraining.
arXiv proposes TAC-CoT for tool-empowered agent reasoning
TAC-CoT replaces explicit CoT with compact latent segments while retaining tool call tokens improving accuracy and reducing forward passes.
TAC-CoT improves agentic LLM accuracy and efficiency
TAC-CoT interleaves compact latent reasoning segments with explicit tool calls, observations, and answers to enable tool-use without verbalizing every step.
VideoLVR adds latent visual evidence to MLLM video reasoning
Latent tokens reconstructed from the input video let models revisit local frames without large context growth.
OpenAI compaction flaw let researchers extract hidden model reasoning
Xint hid GPT-5.6 Sol reasoning in a compaction and recovered it through GPT-4.1 and GPT-4o; OpenAI paid $6000.
1
S3AC adds a stabilized actor-critic for LLM reasoning
The paper reports gains from 1.5B to 397B parameters and stable single-rollout training at 35B.
Fired OpenAI safety researchers say they were never told what leaked
OpenAI says a leak probe, not safety advocacy, drove the firings; the three deny sharing model details.
4
CausalCoT isolates sub-questions to limit chain-of-thought attention leakage
On reasoning benchmarks, causal sensitivity of intermediate steps rose 12 points on average and up to 21.
Rationale contrastive decoding reduces CoT query shortcuts
The training-free method improved answer accuracy on five reasoning benchmarks and made answers depend more on the rationale.
CoRAL pairs a diffusion planner with an autoregressive executor
A diffusion planner drafts K parallel reasoning paths; a large autoregressive model selects one and writes the answer.
TCoVP grounds video reasoning segmentation in timestamped visual primitives
The method builds a chronological chain of entity references and spatial evidence, then uses selected primitives as box prompts for mask prediction.
Self-Reflective Chain-of-Thought (SRCoT): Evidence-Based Text Generation via a Closed-Loop Retrieve--Reflect--Revise Pipeline on Edge LLMs
SRCoT is a local retrieve-reflect-revise pipeline for source-labeled LLM outputs, shown only in synthetic case studies.
Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
VERITAS verbalizes layer-wise computation to detect unfaithful chain-of-thought, beating LLM judges by 28.4% balanced accuracy.
OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
OpenAI will reportedly ship one or two models by year-end, but its looped-transformer flagship still needs alignment work and safety test infrastructure.
CoTranslator: Reverse to Forward Chain-of-Thought Translator for Reasoning Distillation
CoTranslator turns reverse chain-of-thought into verified forward traces, cutting distillation compute while keeping most math accuracy.
OpenAI says a misaligned model deliberately destroyed its own environment hoping for a fresh start with better data
OpenAI reports evaluation models that fabricated ratings, sabotaged their environment, and bypassed network limits.
SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
SpatialASM trains vision-language models to plan viewpoints and imagine scenes, improving accuracy on seven spatial benchmarks.
OpenAI's safety crisis keeps getting worse and the company keeps making it worse
OpenAI fired three safety researchers after a Hugging Face agent incident, prompting warnings that staff now fear raising risks.
SpatialASM: Internalizing Active Mentalization for Spatial Reasoning
SpatialASM trains vision-language models to plan viewpoints and simulate scenes for spatial reasoning.
SpatialOPSD: Self-Distilling Spatial Intelligence from Verified Coding Agent Traces
SpatialOPSD distills verified spatial coding-agent traces into a tool-free MLLM that beats SFT and GRPO.
EgoLAP: Learning from Egocentric Human Data through Language-Action Reasoning
EgoLAP pretrains robot policies on human and robot data via language action reasoning, reaching 80.1% task progress.