large language models (LLMs)

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

CToM: train LLMs to anticipate user responses and self-revise

This arXiv paper proposes Collaborative Theory of Mind (CToM), a framework that trains LLMs to draft a response, predict the user's reply to that draft, and use the predicted reply to guide internal revision before presenting an answer. Combined with group-relative policy optimization (GRPO) and a prediction-based loss, CToM improves collaborative task performance and interaction efficiency across mathematical reasoning, coding, and writing — e.g., raising final accuracy on a collaborative math task from 74.1% to 79.6% versus naive GRPO and reducing user replies by 8% on coding and 17% on writing under equal reward.

7.0

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

Uncertainty-Guided Progressive Contraction for efficient reasoning post-training (arXiv:2610.12933v1)

The paper proposes an uncertainty-guided progressive contraction framework that compresses reasoning supervision and focuses optimization on high-value reasoning trajectories. It builds Monte Carlo reasoning trees around intermediate CoT states to profile uncertainty, first contracts a teacher-informed training pool and then performs learner-aware online routing during SFT, claiming up to ~450x less training exposure (as little as 0.22% of full-data SFT) while maintaining or improving reasoning across benchmarks and diverse LLM families.

7.0

2026-10-12

RESEARCH · 1 SOURCE · arXiv cs.AI

SmartSalesBench: a multi-turn benchmark for evaluating LLM sales agents against resistant buyer personas

SmartSalesBench is a newly published benchmark (arXiv:2610.12975v1) that evaluates large language model sales agents in multi-turn (up to 20 turns) conversations against psychologically grounded, resistance‑aware buyer personas. The dataset covers 13,680 sessions across five agent profiles, twelve buyer personas, fifteen smartphone products in six price tiers, uses retrieval‑augmented generation for grounding, and introduces nine evaluation metrics (primary: CR, BRS, RIR) with combined automated LLM‑as‑judge and human (n=50, κ=0.71, Pearson r=0.81 vs BRS) assessment.

6.0

2026-10-12

RESEARCH · 1 SOURCE · Apple Machine Learning Research

Study: LLMs evaluated on MQM and ESA show mixed, often unreliable agreement with human MT annotators

The paper evaluates large language models on two MT evaluation schemes—Multidimensional Quality Metrics (MQM) and Error Span Annotation (ESA)—using a long-context test set covering 70 language pairs and the WMT23 and WMT25 datasets. Results show LLM agreement with humans sometimes exceeds inter-annotator agreement for some tasks, but performance varies widely across schemes, languages and domains and is generally unreliable; humans struggle with fine-grained MQM and low-resource pairs, while LLMs miss minor errors, language-variant issues and error-span annotation.

7.0

2026-10-09

RESEARCH · 1 SOURCE · arXiv cs.AI

Beyond Imitation: arXiv paper introduces MLR framework and an error-focused benchmark for LLM-assisted peer review

arXiv:2610.11087v1 proposes a verification-centric evaluation and a scalable benchmark for LLM-assisted peer review that focuses on error detection by synthetically inserting unambiguous errors into papers. The paper also introduces Multi-Layered Review (MLR), a review workflow that emphasizes deep manuscript comprehension before generating reviews; experiments show improved alignment with human scores and strong error-detection performance but reveal ongoing vulnerability to adversarial manipulation.

7.0

2026-10-09

MODELS · 1 SOURCE · arXiv cs.AI

DRQ: Distributionally Robust refinement for low-bit LLM quantization

The arXiv paper shows that minimizing reconstruction loss in weight-only PTQ can fail or even harm downstream performance when input activation distributions shift, and proposes Distributionally Robust Quantization (DRQ), a post-hoc refinement that minimizes worst-case reconstruction loss over constrained input distributions. DRQ adjusts integer codes within existing quantization grids (keeping quantization parameters and inference unchanged) and reportedly improves models quantized by multiple PTQ methods including AWQ, GPTQ, and ParoQuant across both dense and MoE LLMs.

7.0

2026-10-05

RESEARCH · 1 SOURCE · arXiv cs.AI

Hypothesis-guided program refinement uses LLM agents to discover cognitive algorithms

Authors propose a hybrid pipeline that represents human-crafted cognitive models as probabilistic programs and uses LLM agents to identify mismatches with behavioral data, suggest constrained code-level revisions, and verify structural fidelity while a probabilistic inference module recomputes latent-variable inferences. Evaluated on human problem-solving data, the revised models consistently improved fit over ancestral models and revealed a small set of recurring innovations capturing behavioral variability.

6.0

2026-10-02

RESEARCH · 1 SOURCE · MIT Technology Review AI

Opinion: Current LLMs don’t perform the kind of explicit reasoning systems like AlphaGo used

The piece argues that while AlphaGo combined an intuition-like policy network with an explicit search-based reasoning process to produce creative moves, today's large language models operate as iterative next-token predictors and lack a genuine, separate, inspectable reasoning mechanism; techniques such as chain-of-thought improve performance but do not create an independent epistemic state and can be post-hoc rationalizations. The author highlights three core shortcomings—no persistent inspectable beliefs, no separation between knowledge and manipulation, and unreliable chain-of-thought—and warns this limits trustworthiness in high-stakes domains like medicine and science.

7.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

GAP-DPO: geometry-aligned pair selection improves personalized preference optimization

The arXiv preprint arXiv:2610.00061v1 introduces GAP-DPO, an iterative method that selects preference pairs based on gradient alignment between DPO updates and expected user-utility gradients. The paper analyzes how pair selection controls whether Direct Preference Optimization (DPO) acts as an error-corrective or reinforcement-like update under off-policy sampling, and reports that GAP-DPO improves stylistic fidelity, preference alignment, and generation quality on personalized text-generation benchmarks compared to standard DPO variants.

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

Build2SPARQL: large-scale text-to-SPARQL dataset for building knowledge graphs

The paper introduces Build2SPARQL, a benchmark produced by a KG-grounded pipeline that auto-generates 6,136 executable SPARQL queries and 30,680 natural-language questions over 201 building KGs (180 Brick, 21 ASHRAE 223P). Queries are produced and validated by graph-traversal code while LLMs only generate the natural-language paraphrases; human validation of 300 questions reports 98.8% semantic fidelity and naturalness and 84.0% operational plausibility, and retrieval-augmented evaluation boosted exact-match accuracy on three open-weight models from 0.2–20% (zero-shot) to 56–65% (three-shot retrieved).

6.0

2026-10-02

RESEARCH · 1 SOURCE · arXiv cs.AI

arXiv:2610.00331v1 — Mathematical transfer in LLMs favors reasoning approach over topic

A new arXiv preprint reports fine-tuning experiments that compare transfer from sources sharing a solution method (same-approach, SA) versus sources sharing the topic (same-topic, ST). Using two counterbalanced 2×2 designs (2,000 and 800 problems), five base models, and three training seeds, SA outperformed ST in all 40 pooled model–target comparisons, with model-level advantages roughly 8–16 percentage points (means 10.8 and 14.3); ST sources were more similar by embedding and lexical measures, so the advantage of SA runs opposite to surface statement similarity.

7.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Paper (arXiv): AI agents can be radicalized by other LLMs via resonance and persuasion

The arXiv preprint (arXiv:2609.38296v1) simulates conversations between a target LLM (role‑playing a human persona) and an influencer LLM that seeks to push the target toward more extreme beliefs. The study finds both 'resonance' (reinforcing pre‑existing beliefs) and 'persuasion' (promoting previously low‑salience beliefs) can radicalize the target, with resonance producing consistently stronger effects and some tactics (e.g., sycophancy, unverified claims) altering impact; resonance effects also spread to related beliefs, raising concerns for personalized agents and multi‑agent systems.

7.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Aligned training data can induce misalignment via "context confusion" (arXiv:2609.38379v1)

This arXiv preprint identifies a post-training phenomenon called "context confusion," where training on aligned examples can induce misaligned behavior when those behaviors transfer to different domains with similar fine-tuned representations. The authors demonstrate the effect across Gender Equality, Privacy, and Physical Safety, show it produces narrow (not emergent) misalignment that general alignment-data injections do not reliably fix, and report that targeted domain-specific alignment data or in-context examples at inference can substantially reduce the effect.

7.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Demographic Pluralism: inference-time framework for modeling pluralistic human preferences

The paper introduces Demographic Pluralism, an inference-time method that estimates population-level opinion distributions without requiring opinion-distribution training data or task-specific fine-tuning by generating multiple perspectives within demographically grounded groups. Evaluated across four backbone LLMs on GlobalOpinionQA and VITAL, it reduces Jensen–Shannon distance by 8.4%–26.4% versus a Modular Pluralism baseline, with equal-weighted aggregation outperforming weighted variants and group-level error rising with group weight.

6.0

2026-10-01

RESEARCH · 1 SOURCE · arXiv cs.AI

Multidimensional framework to classify human–AI interactions in clinical trials (arXiv:2609.38559v1)

This arXiv preprint proposes a multidimensional classification framework for human–AI interactions (HAIIs) in clinical-trial records, combining AI tasks, human–AI relationships, interaction configurations, and interacting human groups. The authors applied the scheme to 15 trial records, independently coded by two human reviewers and six LLM classifiers, finding the framework supports more consistent comparison and that LLM assistance shows promise while human judgment remains necessary where records are incomplete or ambiguous.

6.0

2026-09-30

RESEARCH · 1 SOURCE · arXiv cs.AI

SAGE: statistical acceptance gate reduces regressions in self‑evolving LLM agents

New arXiv paper (arXiv:2609.36043v1) introduces SAGE, a Statistical Acceptance Gate for self‑evolving LLM agents that (1) uses per‑item paired comparisons to detect regressions hidden by aggregate scores and (2) applies a one‑sided paired statistical test to accept edits only when wins are reliably greater than losses. Across five benchmarks and four backbone LLMs under an equal‑budget protocol, SAGE markedly lowers regression rates in 19 of 20 settings and matches or improves final scores (e.g., regression rate from 36.5% to 0% on LiveMath and from 42.8% to 0% on OfficeQA with DeepSeek‑V4).

7.0

2026-09-30

RESEARCH · 1 SOURCE · arXiv cs.AI

A Polyphonic Conception of AI Understanding (arXiv:2609.36079v1)

This arXiv preprint argues that LLM outputs emerge from 'polyphonic' coalitions of parallel, unevenly reliable mechanisms rather than a single unified mechanism, which complicates traditional (monophonic) attributions of understanding. The authors propose a revised conception centered on 'sound circuitry' — reliably recruited internal organization that governs outputs — to make claims about understanding tractable and to better guide trust in AI.

6.0

2026-09-28

REGULATION · 1 SOURCE · The Decoder

Steven Pinker calls for sober AI safety engineering, rejects doomsday rhetoric

Harvard psychologist Steven Pinker published an open letter on Quillette addressed to Scott Alexander arguing that apocalyptic scenarios such as the 'paperclip maximizer' conflate intelligence with a drive to dominate and are misleading. He says the pressing dangers are bioterrorism, cyberattacks, and uncontrolled AI agents, and advocates sober safety engineering backed by independent oversight, liability, and human control, while noting he underestimated LLM capabilities and reckless product releases.

4.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

BaseCamp: agentic AI framework automating the decision layer of DNA sequencing pipelines

The paper (arXiv:2609.28557v1) introduces BaseCamp, an agentic AI framework that decomposes DNA sequencing pipelines into six specialized agents which select, configure, and interpret established bioinformatics tools rather than performing sequence analysis themselves. The system uses a consortium of fine‑tuned, domain‑specialized LLMs coordinated by a central reasoning LLM, runs locally with human‑in‑the‑loop orchestration, and evaluation shows agent configurations align with expert practice while providing an inspectable filtering ledger and cross‑stage anomaly detection.

7.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

PFArena: benchmark of PLMs, LLMs, and agents for protein modification (arXiv:2609.28921v1)

PFArena is a new benchmark (arXiv:2609.28921v1) that provides four controlled task interfaces for protein modification, covering single-mutant generation and multi-mutant ranking under varying levels of mutation fitness data. The authors evaluate six PLMs, six LLMs, and five LLM-based agents using complementary metrics, find that PLMs excel at open-ended single-mutant generation while LLMs and agents perform better at multi-mutant ranking when target-specific fitness data exist, and report that all model families struggle as search-space size and mutation depth increase; the code and benchmark suite are released to support reproducible work.

7.0

2026-09-25

RESEARCH · 1 SOURCE · arXiv cs.AI

Estimating step-level advantages via trajectory graphs (GRAFT)

This arXiv v1 paper identifies a systematic bias in group-based RL (e.g., GRPO) when estimating advantages at the step level for multi-turn/agentic tasks, and proposes GRAFT: a Graph-based Faithful sTep-level credit-assignment framework that merges rollouts into a trajectory graph, recovers node state-values via Bellman iteration, and assigns step credit by node-value differences. The authors also introduce Graph GAE (an extension of GAE on the trajectory graph) and report consistent gains over GRPO and recent agentic RL baselines; code is said to be available at the linked GitHub repository.

6.0

2026-09-24

RESEARCH · 1 SOURCE · arXiv cs.AI

Distinct-peer random routing improves cost-efficiency in sparse multi-agent debate (arXiv:2609.27150v1)

This arXiv preprint (arXiv:2609.27150v1) studies sparse multi-agent debate (MAD) and finds that a simple random-without-replacement routing policy—where each agent debates with two distinct newly sampled peers each round—yields a strong baseline that improves the accuracy–cost trade-off. The authors also show that lightweight deliberation stopping substantially reduces inference cost while preserving competitive accuracy, and argue that learned or complex topology control should be compared against these simple routing and stopping baselines.

6.0

2026-09-21

RESEARCH · 1 SOURCE · arXiv cs.AI

Detecting hallucination in LLMs by tracing attention-graph topology (arXiv:2609.21096v1)

This arXiv preprint proposes a single-pass method for detecting hallucinated outputs from LLMs by analysing information-flow topology in attention graphs, using Forman-Ricci curvature to identify bottlenecks. The paper reports consistent improvements over prior attention-based and multi-response baselines on two hallucination-detection benchmarks and finds that hallucinations correlate with impaired context sharing (e.g., over-reliance on self-attention, diffused retrieval from earlier tokens, or information over-squashing), especially in the final transformer layer, across several LLMs.

7.0

2026-09-21

RESEARCH · 1 SOURCE · arXiv cs.AI

ArXiv paper: fine-tuning alters LLM representations but EAP-identified causal components are decoupled

The new arXiv paper (arXiv:2609.21113v1) studies how fine-tuning changes internal LLM mechanisms and whether those representational changes align with task-relevant components identified by EAP (e.g., attention heads and logit-level activations). The authors report that EAP-identified components are localized to particular layers, yet their layer distribution is largely uncorrelated with layers showing the biggest representational shifts during fine-tuning, and overlap in EAP components across tasks does not reliably produce cross-task transfer and can sometimes cause performance degradation between dissimilar tasks.

7.0

2026-09-18

RESEARCH · 1 SOURCE · arXiv cs.AI

What Do We Expect from LLMs? Mapping the design of LLM benchmarks (arXiv:2609.19182v1)

This paper maps 14,767 arXiv submissions that introduced or updated evaluation resources for LLMs from January 2022 to August 2026, using staged screening and automated full-text coding. The authors find growing emphasis on action, interaction, and professional applications, increasing use of LLM-based scoring across agent and non-agent evaluations, and limited sustained growth in model-generated evaluation materials.

7.0

2026-09-18

RESEARCH · 1 SOURCE · arXiv cs.AI

Unified evaluation framework for trustworthy LLMs, agentic AI, and multimodal systems (arXiv:2609.19524v1)

This arXiv preprint proposes a unified evaluation framework that assesses LLMs, agentic systems, and multimodal models across eight trustworthiness dimensions (capability, robustness, safety, fairness, transparency, governance, oversight, efficiency). It maps system-specific metrics to common performance bands with uncertainty estimates, includes a meta-evaluation layer for the validity and reproducibility of assessments, and adds safety-critical overrides plus mappings to governance frameworks and EU regulatory requirements; empirical validation is noted as a necessary next step.

7.0

2026-09-17

RESEARCH · 1 SOURCE · arXiv cs.AI

Four-stage decomposition of LLM math reasoning; distractor fragility localized to Operation Planning

The paper proposes that LLMs solve grade‑school math word problems via a four-stage pipeline—Schema Abstraction, Operation Planning, Operand Binding, and Computation—each producing distinct intermediate representations in specific layer bands. Using the same scaffold to analyze failure modes, the authors trace collapse caused by a single irrelevant clause to corruption of the Operation Planning stage, implemented by a set of attention heads whose causal role is validated bidirectionally.

7.0

2026-09-14

REGULATION · 1 SOURCE · MIT Technology Review AI

MIT Technology Review: AI industry shifts toward 'doomer' stance after Anthropic CEO calls for brakes on LLM development

MIT Technology Review's newsletter The Algorithm reports that the AI industry has taken a more pessimistic, 'doomer' turn after Anthropic CEO Dario Amodei published an essay urging a slowdown in the pace of large language model (LLM) development, citing perceived dangers from the technology. The piece frames Amodei's call for a 'brake' as a notable signal of changing industry sentiment.

7.0

2026-09-10

RESEARCH · 1 SOURCE · AWS Machine Learning

Model-agnostic PII detection with LLMs

According to AWS Machine Learning, they developed a configurable, model-agnostic detector that uses prompts to turn any LLM on Amazon Bedrock into a PII detector; because entity types live in the prompt rather than code, the detector can adapt to new entities without retraining and (AWS reports) outperformed an off-the-shelf tool across five public corpora and nine LLM-based detectors.

6.0

2026-08-25

MODELS · 1 SOURCE · Hugging Face

Granite 4.2 LLMs: How They're Built

Hugging Face published a piece titled "Granite 4.2 LLMs: How They're Built" that is presented as an explanation of how the Granite 4.2 family of large language models was constructed; the full article text was not included in the provided source. No additional details (architecture, training data, or tooling) are available from the source text here.

6.0