LLMs Can Fix Math Errors But One Token Breaks Them

The authors show that large language models not only store features in superposition but also perform feature‑specific error correction. Across six open‑source models they detect privileged "pure" feature directions that resist perturbations, challenging prevailing assumptions about how LLMs represent and fix internal mistakes.

The TSJ framework simulates months of interaction between LLM‑powered AI companions and developing users, uncovering risks that only appear after roughly 140 conversational turns. Children and emerging adults show heightened vulnerability in trust and emotional dependence, meaning standard short‑term safety tests vastly underestimate real‑world harm.

Researchers define 'cliff tokens', individual tokens that cause a sudden drop in reasoning potential, flipping a correct solution into failure. Across seven models and three benchmarks, removing the first cliff token restores performance to perfect scores, while targeting them with specialized DPO training boosts accuracy up to 6.6 points. This pinpoints a new granularity of LLM fragility.

The paper shows that dense per‑loop cross‑entropy only supervises readout‑exposed variables, leaving recurrent scale unchecked; architectures like RMSNorm hide this scale, causing hidden‑state norms to balloon. Making scale visible or removing it from the loop stabilizes training and improves perplexity.

DeepSWE is an open-source, long-horizon software engineering benchmark that evaluates coding agents on 113 original tasks drawn from 91 active repositories across five languages. All prompts are written from scratch, ensuring no pre-training contamination, and each task includes full test suites to measure real-world bug-fix and feature-add performance.

Researchers combined a vague intuition about rational approximations with the agentic AI system AIM, turning it into a concrete problem and a family of sign‑embedding quantum algorithms for matrix equations. The work shows AI can drive early-stage mathematical insight, not just finish pre‑set proofs, reshaping how quantum algorithm design may progress.

Post‑training INT3/INT4 quantization keeps accuracy but forces reasoning models to generate longer chains of thought, inflating token counts by up to 2×. The extra compute erodes the expected latency benefits, meaning low‑bit inference may not speed up real‑world deployments.

TrustMem adds a verification layer to LLM agents' memory updates, catching omissions, corruptions and hallucinations before they become permanent. The system’s preference‑guided RL boosts memory utility, cutting error rates by up to 79% and raising benchmark F1 scores by 12 points, making long‑term LLM memory far more reliable.

Mistral’s new OCR 4 model delivers structured extraction, text, bounding boxes, block types and confidence scores, in a single‑container deployment covering 170 languages. Independent tests show a 72% win rate over leading OCR systems, making it a strong fit for enterprise search, RAG pipelines, and compliance‑focused self‑hosting.

The Trump administration is urging Meta to submit its latest frontier AI model, Muse Spark, for voluntary security review, the only major U.S. developer yet to cooperate. A government window could expose vulnerabilities before wide release, and Meta’s refusal leaves a critical gap in national‑security oversight.

The paper analyses the new US bill H.R. 238 and Utah’s prescription‑renewal pilot that authorize AI to prescribe drugs, and proposes a ‘clinician’s veto’ architecture requiring calibrated confidence scores, uncertainty type disclosure, and real‑time inferential transparency. A survey of 136 U.S. clinicians shows they would only accept autonomous prescribing if those safeguards are in place, tying liability to system designers.

Gefen replaces AdamW with a single‑precision state store, trimming optimizer memory footprint up to eightfold, roughly 6.5 GiB saved per billion parameters, while preserving AdamW‑level convergence. The change is a two‑line code swap, unlocking bigger models or batch sizes on modest hardware.