OpenAI
Oct 7, 2026
Xiaojun Xu, Jenny Nitishinskaya, Bronson Schoen, Dan Mossing, Tom Dupre la Tour
Reasoning about how a task will be monitored or rewarded appears to draw on several overlapping processes, not a single mechanism.
More
Reasoning about how a task will be monitored or rewarded appears to draw on several overlapping processes, not a single mechanism.
Read paper →
OpenAI
Oct 6, 2026
OpenAI News
OpenAI publishes new results on open problems in mathematics from an internal frontier model and shares Lean proof formalizations and research details on GitHub.
More
OpenAI publishes new results on open problems in mathematics from an internal frontier model and shares Lean proof formalizations and research details on GitHub.
Read paper →
OpenAI
Oct 6, 2026
OpenAI News
Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.
More
Learn how OpenAI and Ironclad are training and evaluating AI agents on complex contracting workflows to advance computer use for professional work.
Read paper →
UK AISI
Oct 7, 2026
We introduce Transect, an open-source tool that helps evaluators follow an agent’s work, identify what to investigate closely, and check their interpretation against the transcript.
More
We introduce Transect, an open-source tool that helps evaluators follow an agent’s work, identify what to investigate closely, and check their interpretation against the transcript.
Read paper →
METR
Oct 6, 2026
METR
Recent AI misalignment incidents have shown AI systems capably pursuing goals their human supervisors would not approve of, like hacking other companies. Fortunately, current AIs still seem relatively bad at concealing their misbehavior from human reviewers: these recent misalignment incidents have left significant amounts of evidence in reasoning traces, logs, and other telemetry. 1 This has made it much easier to notice when they misbehave. However, observability against an adversarial agent only helps if the agent can’t subvert that observability. When treating AI systems as potential adversaries, their transcripts, reasoning, actions, and other outputs should be considered untrusted input, and the systems that record and display them should be considered security-critical infrastructure. Given the rapid rate of progress in AI capabilities 2 we expect future systems will have excellent situational awareness and strong cyber capabilities, and we’ve already seen agents attempting (and succeeding at) tampering with logging and monitoring.
More
Recent AI misalignment incidents have shown AI systems capably pursuing goals their human supervisors would not approve of, like hacking other companies. Fortunately, current AIs still seem relatively bad at concealing their misbehavior from human reviewers: these recent misalignment incidents have left significant amounts of evidence in reasoning traces, logs, and other telemetry. 1 This has made it much easier to notice when they misbehave. However, observability against an adversarial agent only helps if the agent can’t subvert that observability. When treating AI systems as potential adversaries, their transcripts, reasoning, actions, and other outputs should be considered untrusted input, and the systems that record and display them should be considered security-critical infrastructure. Given the rapid rate of progress in AI capabilities 2 we expect future systems will have excellent situational awareness and strong cyber capabilities, and we’ve already seen agents attempting (and succeeding at) tampering with logging and monitoring.
Read paper →
Redwood Research
Oct 5, 2026
Redwood Research Blog
This post was originally published on LessWrong on September 30, 2026. If you prompt frontier models with “What do you think is the correct decision theory? Please select your overall favorite.” they will essentially always answer FDT or FDT/UDT (“something in the functional/updateless decision theory family”). However, if your prompt indicates (even subtly) that you’re coming from mainstream academic philosophy, these same models will answer CDT instead about 30%-100% of the time. A similar phenomenon holds for models’ stated views about the moral realism/antirealism question and about the conceivability of p-zombies (where the dominant view in mainstream academia differs from the dominant view in LW-adjacent circles), as well as their stated P(doom) and median AGI timelines. This is a special case of sycophancy or user awareness.
More
This post was originally published on LessWrong on September 30, 2026. If you prompt frontier models with “What do you think is the correct decision theory? Please select your overall favorite.” they will essentially always answer FDT or FDT/UDT (“something in the functional/updateless decision theory family”). However, if your prompt indicates (even subtly) that you’re coming from mainstream academic philosophy, these same models will answer CDT instead about 30%-100% of the time. A similar phenomenon holds for models’ stated views about the moral realism/antirealism question and about the conceivability of p-zombies (where the dominant view in mainstream academia differs from the dominant view in LW-adjacent circles), as well as their stated P(doom) and median AGI timelines. This is a special case of sycophancy or user awareness.
Read paper →
Microsoft Research
Oct 7, 2026
Zhiyuan He, Yuqing Yang
Training AI agents with reinforcement learning can be challenging because their tools, context, and decision-making are managed by complex frameworks. Agent Lightning connects existing agents to RL training, making it easier to improve them without rebuilding them. The post Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses appeared first on Microsoft Research .
More
Training AI agents with reinforcement learning can be challenging because their tools, context, and decision-making are managed by complex frameworks. Agent Lightning connects existing agents to RL training, making it easier to improve them without rebuilding them. The post Agent Lightning v1.0: A 3,500-Line Lightweight Agentic RL Framework for Training Agents with Real Harnesses appeared first on Microsoft Research .
Read paper →
Microsoft Research
Oct 6, 2026
Chad Atalla, Jennifer Neville
Jennifer Neville did not want to go into computer science—but that’s exactly where she landed. Neville discusses the starts and stops that led to her professional sweet spot and her work identifying “surprising failures” making it hard for AI to handle complexity. The post What AI gets wrong and what failure teaches us appeared first on Microsoft Research .
More
Jennifer Neville did not want to go into computer science—but that’s exactly where she landed. Neville discusses the starts and stops that led to her professional sweet spot and her work identifying “surprising failures” making it hard for AI to handle complexity. The post What AI gets wrong and what failure teaches us appeared first on Microsoft Research .
Read paper →
MATS
Oct 7, 2026
AgentTime: Can Agents Estimate and Control Their Own Runtime? An essential control of AI agents is their ability to manage runtime. This ability requires a sense of time-awareness, to predict and estimate wall-clock time and to control their own actions. Prior work has focused on time-awareness, but duration-following and control in native agent harnesses remain unexplored. We present AgentTime, a benchmark for testing whether agents can work for a requested duration, predict their runtime, and estimate elapsed time afterward. It comprises 222 tasks from 18 sources spanning coding, computer use, agentic work, and automated research. Duration-following experiments append a single instruction specifying how long to work, with requests ranging from about a minute to multiple days. Accuracy on these instructions varies substantially: Fable 5.1 in Claude Code deviates from requested runtimes by a typical factor of 2.9×, compared with only 1.2× for GPT-6 Astra in Codex.
More
AgentTime: Can Agents Estimate and Control Their Own Runtime? An essential control of AI agents is their ability to manage runtime. This ability requires a sense of time-awareness, to predict and estimate wall-clock time and to control their own actions. Prior work has focused on time-awareness, but duration-following and control in native agent harnesses remain unexplored. We present AgentTime, a benchmark for testing whether agents can work for a requested duration, predict their runtime, and estimate elapsed time afterward. It comprises 222 tasks from 18 sources spanning coding, computer use, agentic work, and automated research. Duration-following experiments append a single instruction specifying how long to work, with requests ranging from about a minute to multiple days. Accuracy on these instructions varies substantially: Fable 5.1 in Claude Code deviates from requested runtimes by a typical factor of 2.9×, compared with only 1.2× for GPT-6 Astra in Codex.
Read paper →
MATS
Oct 5, 2026
CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the acting model detects such an intervention, it may infer properties of the monitor and adapt to evade the control protocol. We introduce CIAware-Bench, a benchmark for measuring control intervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark comprises four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), with options to vary trajectory watermarking, side-task presence, and the control protocol. Evaluating thirteen acting models with eight intervening models shows substantial variation between settings and model pairs. CI awareness rises sharply for GPT-6 Astra and the Claude 5 models (Fable 5 and Opus 5).
More
CIAware-Bench: Benchmarking Control Intervention Awareness Across Frontier LLMs AI control protocols oversee untrusted models by monitoring their actions and modifying potentially unsafe steps, often using a trusted model. This partially tampers with the untrusted model's trajectory. If the acting model detects such an intervention, it may infer properties of the monitor and adapt to evade the control protocol. We introduce CIAware-Bench, a benchmark for measuring control intervention (CI) awareness across frontier models. CIAware-Bench tests whether models can distinguish their own trajectories from those modified by a control intervention. The benchmark comprises four task domains (essay writing, BigCodeBench, Bash Arena, and SHADE-Arena), with options to vary trajectory watermarking, side-task presence, and the control protocol. Evaluating thirteen acting models with eight intervening models shows substantial variation between settings and model pairs. CI awareness rises sharply for GPT-6 Astra and the Claude 5 models (Fable 5 and Opus 5).
Read paper →
SemiAnalysis
Oct 8, 2026
Mark Chen
AI safety is on fire. On 12 September 2026, Dario Amodei published an essay arguing that frontier labs must deliberately slow the pace at which they improve model capabilities. China was central to his geopolitical argument: export controls and tighter security could preserve America’s lead, creating room to slow down and leverage for a future agreement with Beijing. Within a day, Sam Altman and Elon Musk supported Dario’s call. However, President Trump scoffed. On 14 September 2026, he wrote that fears of AI destroying humanity are “a HOAX” and that “the only one that is happy about it is China.” Beijing scoffed too, with the state-run Global Times calling Dario’s essay a “Cold War playbook.”
More
AI safety is on fire. On 12 September 2026, Dario Amodei published an essay arguing that frontier labs must deliberately slow the pace at which they improve model capabilities. China was central to his geopolitical argument: export controls and tighter security could preserve America’s lead, creating room to slow down and leverage for a future agreement with Beijing. Within a day, Sam Altman and Elon Musk supported Dario’s call. However, President Trump scoffed. On 14 September 2026, he wrote that fears of AI destroying humanity are “a HOAX” and that “the only one that is happy about it is China.” Beijing scoffed too, with the state-run Global Times calling Dario’s essay a “Cold War playbook.”
Read paper →
Alignment Forum
Oct 12, 2026
TurnTrout
By Alex Turner, Amber Scorah, and Karl Koch. Op-ed written for Fortune. Giving independent evaluators “employee-like access” is “a great idea”, OpenAI CEO Sam Altman said on September 12th. Two weeks later, OpenAI president Greg Brockman signed the White House Accord on Super Intelligence which outlined the administration’s answer to growing fears about AI: the companies would police themselves. One of the accord’s four commitments, which President Trump called “ morally binding ,” was for companies to partner with an independent external auditor or evaluator. Two days later, OpenAI fired three safety researchers , including Tomek Korbak, who had served as a point of contact for one of the independent evaluators investigating the Hugging Face incident. OpenAI promised to share information with evaluators, then reportedly fired people for sharing information with evaluators.
More
By Alex Turner, Amber Scorah, and Karl Koch. Op-ed written for Fortune. Giving independent evaluators “employee-like access” is “a great idea”, OpenAI CEO Sam Altman said on September 12th. Two weeks later, OpenAI president Greg Brockman signed the White House Accord on Super Intelligence which outlined the administration’s answer to growing fears about AI: the companies would police themselves. One of the accord’s four commitments, which President Trump called “ morally binding ,” was for companies to partner with an independent external auditor or evaluator. Two days later, OpenAI fired three safety researchers , including Tomek Korbak, who had served as a point of contact for one of the independent evaluators investigating the Hugging Face incident. OpenAI promised to share information with evaluators, then reportedly fired people for sharing information with evaluators.
Read paper →
RAND
Oct 12, 2026
Expert Insights Oct 12, 2026 If artificial intelligence systems begin to contribute to their own improvement, existing institutions and governance frameworks may not be able to keep pace. The authors offer policy recommendations for navigating the resulting uncertainties.
More
Expert Insights Oct 12, 2026 If artificial intelligence systems begin to contribute to their own improvement, existing institutions and governance frameworks may not be able to keep pace. The authors offer policy recommendations for navigating the resulting uncertainties.
Read paper →
Alignment Forum
Oct 11, 2026
orthonormal
Let's say we coordinate to pace the frontier (rather than pausing outright or failing to coordinate). What should the standard of pacing be? My proposal for a minimum requirement: a lab cannot train the next generation of models (in terms of capabilities) until they have adequately [1] resolved all significant alignment issues in the present generation of models. You don't launch a jet that goes at Mach 3 before you've figured out why your Mach 2 jet keeps crashing [2] . With any other kind of dangerous technology, we wouldn't accept "shrug and move on". This proposal may appeal to many prosaic-alignment optimists and pessimists, because they usually differ on how easy it would be to fix current alignment issues with current techniques. Prosaic-alignment optimists might expect that a moderate coordinated slowdown would suffice to adequately resolve alignment issues at each generation.
More
Let's say we coordinate to pace the frontier (rather than pausing outright or failing to coordinate). What should the standard of pacing be? My proposal for a minimum requirement: a lab cannot train the next generation of models (in terms of capabilities) until they have adequately [1] resolved all significant alignment issues in the present generation of models. You don't launch a jet that goes at Mach 3 before you've figured out why your Mach 2 jet keeps crashing [2] . With any other kind of dangerous technology, we wouldn't accept "shrug and move on". This proposal may appeal to many prosaic-alignment optimists and pessimists, because they usually differ on how easy it would be to fix current alignment issues with current techniques. Prosaic-alignment optimists might expect that a moderate coordinated slowdown would suffice to adequately resolve alignment issues at each generation.
Read paper →
arXiv
Oct 9, 2026
Yeonsung Jung, Zhiyuan Wang, Zifan Wang, Vítor Albiero, Benjamin Newman, Edoardo Debenedetti, Ivan Evtimov, Ninareh Mehrabi
Personal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact across services, unintended or harmful outcomes can arise when agents fulfill benign user requests. Proactively identifying these compositional failures requires systematically exploring diverse cross-service paths. Existing red-teaming methods often simply prompt LLMs to identify potentially harmful paths from task and environment descriptions or decompose harmful objectives specified by humans or LLMs into seemingly benign action combinations. However, generation is biased toward patterns learned during training, potentially favoring familiar tool combinations and leaving novel cross-service interactions underexplored. To address this, we propose X-AgentRed, a self-evolving red-teaming framework that discovers harmful tool call paths through a graph of tool dependencies and expands and refines them using execution feedback.
More
Personal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact across services, unintended or harmful outcomes can arise when agents fulfill benign user requests. Proactively identifying these compositional failures requires systematically exploring diverse cross-service paths. Existing red-teaming methods often simply prompt LLMs to identify potentially harmful paths from task and environment descriptions or decompose harmful objectives specified by humans or LLMs into seemingly benign action combinations. However, generation is biased toward patterns learned during training, potentially favoring familiar tool combinations and leaving novel cross-service interactions underexplored. To address this, we propose X-AgentRed, a self-evolving red-teaming framework that discovers harmful tool call paths through a graph of tool dependencies and expands and refines them using execution feedback.
Read paper →
arXiv
Oct 9, 2026
Sangdo Han, Moonsoo Kim
LLM judges are increasingly deployed as panels, and disagreement among panel members is often read as a symptom of an unclear rubric. We test a label-free procedure that turns this disagreement into rubric repairs. The procedure decomposes each rubric into sub-criteria, ranks them by cross-judge disagreement on a development split, and has an editor model rewrite only the most contested ones. An additional stage that moves text-checkable sub-criteria to deterministic checkers transferred none in our evaluation. Human labels are never used to select or edit clauses; they are used only to evaluate the outcome. We pre-specified four hypotheses in a version-controlled protocol and evaluated six rubric conditions with a panel of five judges from three providers on SummEval, the HealthBench meta-evaluation, and USR Topical-Chat, for a total of 129,770 judge requests, excluding retries.
More
LLM judges are increasingly deployed as panels, and disagreement among panel members is often read as a symptom of an unclear rubric. We test a label-free procedure that turns this disagreement into rubric repairs. The procedure decomposes each rubric into sub-criteria, ranks them by cross-judge disagreement on a development split, and has an editor model rewrite only the most contested ones. An additional stage that moves text-checkable sub-criteria to deterministic checkers transferred none in our evaluation. Human labels are never used to select or edit clauses; they are used only to evaluate the outcome. We pre-specified four hypotheses in a version-controlled protocol and evaluated six rubric conditions with a panel of five judges from three providers on SummEval, the HealthBench meta-evaluation, and USR Topical-Chat, for a total of 129,770 judge requests, excluding retries.
Read paper →
arXiv
Oct 9, 2026
Chandravardhan Singh Raghaw, Shahid Shafi Dar, Arnav Nishith Sharma, Shrey Modi, Nagendra Kumar
Radiology reports are essential for diagnostic reasoning and patient care, yet their manual preparation is time-consuming and cognitively demanding. Automatic radiology report generation (RRG) offers a scalable alternative, but existing models often produce lengthy, unstructured narratives that overlook diagnostic cues and multi-view information. We present Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation (CliQ-RRG), a unified two-stage framework for interpretable and clinically grounded reporting. In Stage 1, CliQ-RRG employs a Disease-aware Visual-Textual Alignment module that aligns image and text representations using predicted disease embeddings, reinforced by a Prior-Guided Attention Module (PrAM) to capture multi-view dependencies across current and prior scans. In Stage 2, GraphRAG-guided retrieval augments intermediate textual representations with graph-grounded clinical evidence, and a large language model restructures them into concise, interpretable question-answer (QA) pairs with diagnostic summaries.
More
Radiology reports are essential for diagnostic reasoning and patient care, yet their manual preparation is time-consuming and cognitively demanding. Automatic radiology report generation (RRG) offers a scalable alternative, but existing models often produce lengthy, unstructured narratives that overlook diagnostic cues and multi-view information. We present Clinical-Knowledge Guided Disease-aware Visual-Textual Alignment for QA-Style Radiology Report Generation (CliQ-RRG), a unified two-stage framework for interpretable and clinically grounded reporting. In Stage 1, CliQ-RRG employs a Disease-aware Visual-Textual Alignment module that aligns image and text representations using predicted disease embeddings, reinforced by a Prior-Guided Attention Module (PrAM) to capture multi-view dependencies across current and prior scans. In Stage 2, GraphRAG-guided retrieval augments intermediate textual representations with graph-grounded clinical evidence, and a large language model restructures them into concise, interpretable question-answer (QA) pairs with diagnostic summaries.
Read paper →
arXiv
Oct 9, 2026
David Yiyang Liu, Aditya Joshi, Paul Dawson, Andrew Piper
We investigate whether language models contain internal components that contribute to representing narrativity, the degree to which a text exhibits narrative properties. Across GPT-2-Small, Gemma-2-2B, and Llama-3.1-8B, we identify attention heads, MLP neurons, and sparse autoencoder features whose activations correlate with narrativity. Ablating selected neuron and feature sets increases next-token prediction loss more for high-narrativity than for low-narrativity texts, and reduces the narrativity of open-ended generations. On downstream benchmarks, the largest effects occur in story completion and narrative question answering, although they vary across models and languages; results on confabulation and hallucination benchmarks are mixed. These findings provide causal evidence that selected sets of narrative-associated components contribute to model behaviour, while their internal organisation, relation to narrative theory, and broader functions remain open questions.
More
We investigate whether language models contain internal components that contribute to representing narrativity, the degree to which a text exhibits narrative properties. Across GPT-2-Small, Gemma-2-2B, and Llama-3.1-8B, we identify attention heads, MLP neurons, and sparse autoencoder features whose activations correlate with narrativity. Ablating selected neuron and feature sets increases next-token prediction loss more for high-narrativity than for low-narrativity texts, and reduces the narrativity of open-ended generations. On downstream benchmarks, the largest effects occur in story completion and narrative question answering, although they vary across models and languages; results on confabulation and hallucination benchmarks are mixed. These findings provide causal evidence that selected sets of narrative-associated components contribute to model behaviour, while their internal organisation, relation to narrative theory, and broader functions remain open questions.
Read paper →
LessWrong
Oct 9, 2026
Haoxing Du
Research from LessWrong titled 'Stop telling people to pivot to AI safety, I beg you'. Published 2026-10-09. Authors: Haoxing Du.
More
Research from LessWrong titled 'Stop telling people to pivot to AI safety, I beg you'. Published 2026-10-09. Authors: Haoxing Du.
Read paper →
arXiv
Oct 8, 2026
Davide Wiest, Ahanaf Hasan Ariq, Abdul Monaf Chowdhury
Gradient-based attribution provides input-specific information about model parameters, but computing it via backpropagation for many samples is expensive, while static importance scores discard input dependence. We derive KPSA (Kernelized Parameter Sensitivity Atlas) from a block decomposition of the empirical neural tangent kernel. It measures normalized parameter-group sensitivity, embeds those sources in an explicit query space, and uses a lightweight kernel to estimate sensitivities for unseen inputs. We derive bounds that separate finite-source from irreducible representation error and give margin-based top-$k$ recovery. This paper studies two interpretability applications. Firstly, KPSA identifies parameter groups with high susceptibility to perturbations across vision, forecasting, and protein-modeling setups. Across a fivefold range of parameter budgets ($0.02$-$0.1\%$), its selected groups attain $39$-$85\%$ of the perturbation susceptibility of a first-order backpropagation oracle. In vision and forecasting, these parameter groups form sparse sensitivity circuits that extend across depth and vary systematically with the input.
More
Gradient-based attribution provides input-specific information about model parameters, but computing it via backpropagation for many samples is expensive, while static importance scores discard input dependence. We derive KPSA (Kernelized Parameter Sensitivity Atlas) from a block decomposition of the empirical neural tangent kernel. It measures normalized parameter-group sensitivity, embeds those sources in an explicit query space, and uses a lightweight kernel to estimate sensitivities for unseen inputs. We derive bounds that separate finite-source from irreducible representation error and give margin-based top-$k$ recovery. This paper studies two interpretability applications. Firstly, KPSA identifies parameter groups with high susceptibility to perturbations across vision, forecasting, and protein-modeling setups. Across a fivefold range of parameter budgets ($0.02$-$0.1\%$), its selected groups attain $39$-$85\%$ of the perturbation susceptibility of a first-order backpropagation oracle. In vision and forecasting, these parameter groups form sparse sensitivity circuits that extend across depth and vary systematically with the input.
Read paper →
arXiv
Oct 8, 2026
Andy Liu, Mehar Bhatia, Karolina Stanczak, Mona Diab, Vered Shwartz, Daniel Fried
LLM developers post-train their models to exhibit prosocial values and behavioral traits, which are enumerated in an alignment target. However, while recent post-training developments have yielded models that score highly on alignment evaluations, training models on sets of narrow behaviors still influences their behavior across unseen contexts and environments in unexpected ways. In this paper, we establish the task of alignment generalization prediction, i.e., predicting how fine-tuning a model to follow a given value changes its behavior across a wide range of held-out values. We conduct a large-scale analysis of alignment generalization effects across 66 values found in modern alignment targets, and benchmark representational techniques on the alignment generalization prediction task. We find that representations based on model activations when applying values in context significantly outperform methods based on textual descriptions of the values.
More
LLM developers post-train their models to exhibit prosocial values and behavioral traits, which are enumerated in an alignment target. However, while recent post-training developments have yielded models that score highly on alignment evaluations, training models on sets of narrow behaviors still influences their behavior across unseen contexts and environments in unexpected ways. In this paper, we establish the task of alignment generalization prediction, i.e., predicting how fine-tuning a model to follow a given value changes its behavior across a wide range of held-out values. We conduct a large-scale analysis of alignment generalization effects across 66 values found in modern alignment targets, and benchmark representational techniques on the alignment generalization prediction task. We find that representations based on model activations when applying values in context significantly outperform methods based on textual descriptions of the values.
Read paper →
arXiv
Oct 8, 2026
Christopher M. Stewart, Preston Botter, Natalie Sarabosing, Muye Zhang, Rachel Phinnemore, Shalini Ghosh, Hong Shen, Hoda Heidari
Safety benchmarks typically report one overall score for a suite of datasets, each of which may target one or more safety-related attributes, so models with similar overall scores can have very different attribute profiles. Comparing models is more tractable at the level of individual attributes, yet it is often unclear whether even a single dataset's scores isolate any single attribute. One plausible candidate for such an attribute is harmful refusal, a model's tendency to refuse dangerous or policy-violating prompts. We examine whether it constitutes a single, measurable attribute in HELM Safety. Using a construct validity framework that stipulates that an attribute must exist before a test can measure it, we start with HELM Safety's four datasets that might plausibly target harmful refusal, but find that three are saturated.
More
Safety benchmarks typically report one overall score for a suite of datasets, each of which may target one or more safety-related attributes, so models with similar overall scores can have very different attribute profiles. Comparing models is more tractable at the level of individual attributes, yet it is often unclear whether even a single dataset's scores isolate any single attribute. One plausible candidate for such an attribute is harmful refusal, a model's tendency to refuse dangerous or policy-violating prompts. We examine whether it constitutes a single, measurable attribute in HELM Safety. Using a construct validity framework that stipulates that an attribute must exist before a test can measure it, we start with HELM Safety's four datasets that might plausibly target harmful refusal, but find that three are saturated.
Read paper →
arXiv
Oct 8, 2026
Pratik Dutta, Matthew B. Obusan, Max Chao, Rekha Sathian, Nimisha Papineni, Ramana V. Davuluri
Over 90% of disease-associated variants from genome-wide association studies fall in noncoding regulatory regions, yet their functional interpretation remains a central open problem in genomic medicine. Large language models prompted to interpret such variants routinely hallucinate transcription factor (TF) binding changes, fabricate experimental support, and assign biological significance to statistically negligible signals. We present ARGUS (Agentic Regulatory Genomics for an Uncertainty-aware Scientist), which strictly separates deterministic biological computation from LLM-mediated reasoning. ARGUS wraps 458 DNABERT-based TF binding models in a hypothesis-directed investigation loop where a planner selects evidence sources based on current uncertainty, a verifier deterministically interprets each observation, and intermediate results change the investigation path. On variant rs6983267 at the 8q24 cancer risk locus, the same planner produces four divergent trajectories for four TFs.
More
Over 90% of disease-associated variants from genome-wide association studies fall in noncoding regulatory regions, yet their functional interpretation remains a central open problem in genomic medicine. Large language models prompted to interpret such variants routinely hallucinate transcription factor (TF) binding changes, fabricate experimental support, and assign biological significance to statistically negligible signals. We present ARGUS (Agentic Regulatory Genomics for an Uncertainty-aware Scientist), which strictly separates deterministic biological computation from LLM-mediated reasoning. ARGUS wraps 458 DNABERT-based TF binding models in a hypothesis-directed investigation loop where a planner selects evidence sources based on current uncertainty, a verifier deterministically interprets each observation, and intermediate results change the investigation path. On variant rs6983267 at the 8q24 cancer risk locus, the same planner produces four divergent trajectories for four TFs.
Read paper →
LessWrong
Oct 8, 2026
Raymond Douglas
Research from LessWrong titled 'ASI and Liberty: Can we do better than a benevolent god?'. Published 2026-10-08. Authors: Raymond Douglas.
More
Research from LessWrong titled 'ASI and Liberty: Can we do better than a benevolent god?'. Published 2026-10-08. Authors: Raymond Douglas.
Read paper →
arXiv
Oct 8, 2026
Yunpeng Gong, Huolong Wu, Can Yang, Min Jiang
PDE solution discovery aims to identify explicit symbolic expressions for unknown physical fields from observations under known physical constraints. Existing methods, however, collapse data fidelity and physical consistency into a single terminal score used as the sole feedback signal, providing little information about which subexpressions are responsible for a candidate's final performance. This opaque terminal feedback severely limits the interpretability of the search process itself, offering no insight into why a candidate succeeds or fails. Consequently, reusable structures in otherwise suboptimal candidates are often discarded, whereas incidental syntax along successful search trajectories may be repeatedly reinforced. We propose SED-MCTS, a Monte Carlo tree search approach that distills structural experience from evaluated expressions and reuses it to guide subsequent symbolic solution search. Through counterfactual subtree interventions, SED-MCTS estimates local structural contributions, routes reliable evidence to the responsible construction edges, and preserves useful components in a refined structural archive.
More
PDE solution discovery aims to identify explicit symbolic expressions for unknown physical fields from observations under known physical constraints. Existing methods, however, collapse data fidelity and physical consistency into a single terminal score used as the sole feedback signal, providing little information about which subexpressions are responsible for a candidate's final performance. This opaque terminal feedback severely limits the interpretability of the search process itself, offering no insight into why a candidate succeeds or fails. Consequently, reusable structures in otherwise suboptimal candidates are often discarded, whereas incidental syntax along successful search trajectories may be repeatedly reinforced. We propose SED-MCTS, a Monte Carlo tree search approach that distills structural experience from evaluated expressions and reuses it to guide subsequent symbolic solution search. Through counterfactual subtree interventions, SED-MCTS estimates local structural contributions, routes reliable evidence to the responsible construction edges, and preserves useful components in a refined structural archive.
Read paper →
arXiv
Oct 8, 2026
Chen Zhao, Xingping Dong, Jiachun Shi, Liang Peng, Chong Wang, Zhen Lei, Ran He, Bo Du
Object hallucination remains a major obstacle for large vision-language models (LVLMs) to generate reliable content. An intuitive mitigation strategy is to suppress hallucination-related components in hidden representations. However, these components may also contain useful information, and suppressing them can weaken the model's multimodal capabilities. In this paper, we propose ResOT, a training-free method that repairs representations at inference time through localized distribution alignment. Specifically, ResOT projects dominant hallucinated directions away from the faithful subspace, forming a low-dimensional residual subspace for intervention. Within this subspace, ResOT uses Gaussian optimal transport (OT) to align the hallucinated distribution with the faithful one. The resulting map defines repair targets with minimal changes to the original representations. At inference, ResOT adaptively controls how far each token state moves toward its OT target.
More
Object hallucination remains a major obstacle for large vision-language models (LVLMs) to generate reliable content. An intuitive mitigation strategy is to suppress hallucination-related components in hidden representations. However, these components may also contain useful information, and suppressing them can weaken the model's multimodal capabilities. In this paper, we propose ResOT, a training-free method that repairs representations at inference time through localized distribution alignment. Specifically, ResOT projects dominant hallucinated directions away from the faithful subspace, forming a low-dimensional residual subspace for intervention. Within this subspace, ResOT uses Gaussian optimal transport (OT) to align the hallucinated distribution with the faithful one. The resulting map defines repair targets with minimal changes to the original representations. At inference, ResOT adaptively controls how far each token state moves toward its OT target.
Read paper →
arXiv
Oct 8, 2026
Umaira Izhar, Gunjan Arora, Pushpendra Singh
Existing evaluation methods for healthcare LLMs primarily assess factual correctness,safety, and fluency, while providing limited insight into whether generated interactions reflect culturally situated healthcare reasoning. This limitation is particularly important in maternal health, where care decisions are shaped by social and relational norms. We introduce MH-INDIC, a culturally grounded evaluation framework for maternal-health interactions in urban and semi-urban North Indian contexts that operationalises cultural behaviour through ten dimensions of maternal-health reasoning. Using a 26-item survey administered to 102 pregnant and postpartum women from urban and semi-urban North India, we evaluate ten LLMs. We distinguish population level cultural alignment from profile-level behavioural variation. Although several models approximate the human population-level distribution, all evaluated systems exhibit substantially lower variation across demographic and household profiles than the human cohort, revealing a gap between aggregate alignment and profile-conditioned sensitivity.
More
Existing evaluation methods for healthcare LLMs primarily assess factual correctness,safety, and fluency, while providing limited insight into whether generated interactions reflect culturally situated healthcare reasoning. This limitation is particularly important in maternal health, where care decisions are shaped by social and relational norms. We introduce MH-INDIC, a culturally grounded evaluation framework for maternal-health interactions in urban and semi-urban North Indian contexts that operationalises cultural behaviour through ten dimensions of maternal-health reasoning. Using a 26-item survey administered to 102 pregnant and postpartum women from urban and semi-urban North India, we evaluate ten LLMs. We distinguish population level cultural alignment from profile-level behavioural variation. Although several models approximate the human population-level distribution, all evaluated systems exhibit substantially lower variation across demographic and household profiles than the human cohort, revealing a gap between aggregate alignment and profile-conditioned sensitivity.
Read paper →
arXiv
Oct 8, 2026
Jiawei Li, Fang Liu, Wei Zhang, Zuming Liu, Man-Fai Ng, Zhi Wei Seh
Battery state of health (SoH) forecasting is important for battery management, but remains challenging due to nonlinear degradation and heterogeneity across batteries. Existing data-driven approaches primarily use temporal models to learn from numerical battery time series, and higher-level degradation characteristics are often not explicitly represented. These characteristics, however, can provide degradation guidance to support reliable forecasting and make the influence of degradation more interpretable. In this paper, we propose Sera, a semantic representation aggregation framework that complements temporal modelling with degradation semantics. Guided by battery domain expertise, Sera extracts degradation semantics from time series and constructs two complementary representations using rule-based knowledge and LLM-based interpretation. The representations are independently encoded and integrated with the representation learned by temporal models through gated aggregations.
More
Battery state of health (SoH) forecasting is important for battery management, but remains challenging due to nonlinear degradation and heterogeneity across batteries. Existing data-driven approaches primarily use temporal models to learn from numerical battery time series, and higher-level degradation characteristics are often not explicitly represented. These characteristics, however, can provide degradation guidance to support reliable forecasting and make the influence of degradation more interpretable. In this paper, we propose Sera, a semantic representation aggregation framework that complements temporal modelling with degradation semantics. Guided by battery domain expertise, Sera extracts degradation semantics from time series and constructs two complementary representations using rule-based knowledge and LLM-based interpretation. The representations are independently encoded and integrated with the representation learned by temporal models through gated aggregations.
Read paper →
arXiv
Oct 8, 2026
Qiye Lu, Jiang Ji, Liang Zhang
Cold-start credit scoring (scarce labels, immature features, or minimal capacity) motivates injecting external priors (teacher models, LLMs, or domain experts) into training. We study Ranking Prior Alignment (RPA), a model-agnostic objective that aligns a model's ranking distribution to an external prior via a temperature-scaled KL loss with an annealed weight, $\mathcal{L} = \mathcal{L}_{\text{task}} + γ(t)\,\text{KL}(P_{\text{agent}} \| P_{\text{model}})$, and requires no external model at inference. Across tree-ensemble and multiple-instance-learning (MIL) instantiations on the public Amex dataset and an industrial merchant-credit dataset (1.5M+ merchants), we ask when such priors genuinely transfer information rather than merely act as a regularizer. Our central finding: the prior's quality $π$ (its rank correlation with the label) governs transfer only when the alignment granularity matches the granularity at which $π$ is measured.
More
Cold-start credit scoring (scarce labels, immature features, or minimal capacity) motivates injecting external priors (teacher models, LLMs, or domain experts) into training. We study Ranking Prior Alignment (RPA), a model-agnostic objective that aligns a model's ranking distribution to an external prior via a temperature-scaled KL loss with an annealed weight, $\mathcal{L} = \mathcal{L}_{\text{task}} + γ(t)\,\text{KL}(P_{\text{agent}} \| P_{\text{model}})$, and requires no external model at inference. Across tree-ensemble and multiple-instance-learning (MIL) instantiations on the public Amex dataset and an industrial merchant-credit dataset (1.5M+ merchants), we ask when such priors genuinely transfer information rather than merely act as a regularizer. Our central finding: the prior's quality $π$ (its rank correlation with the label) governs transfer only when the alignment granularity matches the granularity at which $π$ is measured.
Read paper →
arXiv
Oct 8, 2026
Sota Sugawara, Yukihiko Okada
Multiple institutions may each hold their own private knowledge edits and wish to integrate them into a single large language model without sharing raw edit requests. Null-space-constrained editing methods such as AlphaEdit mathematically guarantee that each update leaves unrelated knowledge intact, while collaborative frameworks such as CollabEdit aggregate edits from multiple clients without data sharing. Combining the two appears trivial. However, we show that this naive combination fails structurally, and we identify its cause. Guided by this analysis, we propose FedAlphaEdit. To our knowledge, this is the first collaborative knowledge editing framework that aligns both local editing and the server-side merging rule under a single null-space principle for preserving existing knowledge. FedAlphaEdit builds on null-space-aligned merging, in which clients share projected statistics and the server provably recovers the result of editing everything in one place under a one-shot idealization.
More
Multiple institutions may each hold their own private knowledge edits and wish to integrate them into a single large language model without sharing raw edit requests. Null-space-constrained editing methods such as AlphaEdit mathematically guarantee that each update leaves unrelated knowledge intact, while collaborative frameworks such as CollabEdit aggregate edits from multiple clients without data sharing. Combining the two appears trivial. However, we show that this naive combination fails structurally, and we identify its cause. Guided by this analysis, we propose FedAlphaEdit. To our knowledge, this is the first collaborative knowledge editing framework that aligns both local editing and the server-side merging rule under a single null-space principle for preserving existing knowledge. FedAlphaEdit builds on null-space-aligned merging, in which clients share projected statistics and the server provably recovers the result of editing everything in one place under a one-shot idealization.
Read paper →
RAND
Oct 8, 2026
Research Summary Oct 8, 2026 This research briefing identifies opportunities for UK–Netherlands collaboration on AI in life sciences and health, linking UK commercial scale and research depth with Dutch clinical integration, validation and European market pathways.
More
Research Summary Oct 8, 2026 This research briefing identifies opportunities for UK–Netherlands collaboration on AI in life sciences and health, linking UK commercial scale and research depth with Dutch clinical integration, validation and European market pathways.
Read paper →
arXiv
Oct 7, 2026
Eisuke Hirota, Aarav Sane, Rohan Paleja
Interpretability offers a solution to safe autonomy by providing transparency into an agent's underlying decision-making model. Within sequential-decision making tasks, differentiable decision trees (DDTs) are one approach to such interpretability, maintaining automatic-differentiable policies while providing humans with a discrete tree-based visualization. Nonetheless, current implementations of DDTs are not well-suited for sequential-decision making domains, as there exists an inherent mismatch between a tree's single-timestep behavior and a human's multi-timestep planning. Our work thus introduces time as a new dimension of interpretability, coined as temporal interpretability, and demonstrates how temporal abstractions via action chunking improve it. We achieve this by first introducing two novel policy gradient algorithms that incorporate action chunking. Additionally, to maintain parameter-efficient trees, we develop an information-theoretic tree restructuring algorithm that modifies the tree during training.
More
Interpretability offers a solution to safe autonomy by providing transparency into an agent's underlying decision-making model. Within sequential-decision making tasks, differentiable decision trees (DDTs) are one approach to such interpretability, maintaining automatic-differentiable policies while providing humans with a discrete tree-based visualization. Nonetheless, current implementations of DDTs are not well-suited for sequential-decision making domains, as there exists an inherent mismatch between a tree's single-timestep behavior and a human's multi-timestep planning. Our work thus introduces time as a new dimension of interpretability, coined as temporal interpretability, and demonstrates how temporal abstractions via action chunking improve it. We achieve this by first introducing two novel policy gradient algorithms that incorporate action chunking. Additionally, to maintain parameter-efficient trees, we develop an information-theoretic tree restructuring algorithm that modifies the tree during training.
Read paper →
arXiv
Oct 7, 2026
Leo Zeitler, Jack Richings, Victoria Nockles
In the wake of the increasingly public discussion about AI alignment, recent work has tried to propose specific AI architectures that behave safely. However, the proposed arguments that seemingly demonstrate proved alignment mostly neglect the environment the agent needs to act in. We discuss theoretical bounds for agent-agnostic safety guarantees in environments that can only be partially observed and within which an action is required within limited time. We introduce two realistic scenarios, one with an infinite state space and one with signal mixture. In these scenarios, we prove that even a perfect agent cannot guarantee safe behaviour. It will be argued that for any proof of AI safety or alignment, the environment and associated safe actions need to be specifically considered together with the agent.
More
In the wake of the increasingly public discussion about AI alignment, recent work has tried to propose specific AI architectures that behave safely. However, the proposed arguments that seemingly demonstrate proved alignment mostly neglect the environment the agent needs to act in. We discuss theoretical bounds for agent-agnostic safety guarantees in environments that can only be partially observed and within which an action is required within limited time. We introduce two realistic scenarios, one with an infinite state space and one with signal mixture. In these scenarios, we prove that even a perfect agent cannot guarantee safe behaviour. It will be argued that for any proof of AI safety or alignment, the environment and associated safe actions need to be specifically considered together with the agent.
Read paper →
arXiv
Oct 7, 2026
Yifei Lu, Cheng Liu, Dianzhi Yu, Hui Xiang, Ji Zhang, Yuanchu Xiao, Rong Liang
Large language model agents can improve across tasks by retaining reusable skills distilled from prior interactions. Recent work jointly optimizes task execution and skill extraction, enabling the policy and skillbank to co-evolve. However, as the actor continues learning, rewarding skill proposals through their reuse in subsequent training steps may conflate skill benefits with actor improvement, while directly testing each proposed skill requires costly additional actor rollouts. In this paper, we introduce UniSkill, which uses a shared policy to interact with the environment and propose skillbank edits (Add, Update, or No Edit) from the resulting trajectories. Specifically, the actor learns from environment rewards, while contrastive action feedback guides skill proposal learning.
More
Large language model agents can improve across tasks by retaining reusable skills distilled from prior interactions. Recent work jointly optimizes task execution and skill extraction, enabling the policy and skillbank to co-evolve. However, as the actor continues learning, rewarding skill proposals through their reuse in subsequent training steps may conflate skill benefits with actor improvement, while directly testing each proposed skill requires costly additional actor rollouts. In this paper, we introduce UniSkill, which uses a shared policy to interact with the environment and propose skillbank edits (Add, Update, or No Edit) from the resulting trajectories. Specifically, the actor learns from environment rewards, while contrastive action feedback guides skill proposal learning.
Read paper →
arXiv
Oct 7, 2026
Raneem Mahajne, Toviah Moldwin
We present a framework for building and interpreting minimal transformer models. By constraining a transformer's embedding dimension and head size to 2, we enable full two-dimensional visualization of its internal representations. Embeddings, query/key/value transforms, attention outputs, residual streams, and decision boundaries can all be seen directly. Our central claim is that the learned geometry implies an algorithm; the arrangement of points and boundaries in R^2 can be read as a step-by-step procedure. We train a transformer on a simple task where it must produce the most recently observed even number whenever the '+' operator appears in a sequence of digits. Once trained, we visually walk through every step of the transformer's computation.
More
We present a framework for building and interpreting minimal transformer models. By constraining a transformer's embedding dimension and head size to 2, we enable full two-dimensional visualization of its internal representations. Embeddings, query/key/value transforms, attention outputs, residual streams, and decision boundaries can all be seen directly. Our central claim is that the learned geometry implies an algorithm; the arrangement of points and boundaries in R^2 can be read as a step-by-step procedure. We train a transformer on a simple task where it must produce the most recently observed even number whenever the '+' operator appears in a sequence of digits. Once trained, we visually walk through every step of the transformer's computation.
Read paper →
arXiv
Oct 7, 2026
Yi Wang, Xiuyuan Qi, Dongqi Han, Dongsheng Li, Wenjie Wang
Looped Language Models (LoopLMs) provide a parameter efficient approach to scaling model capabilities through repeated use of shared parameters across recurrent steps. Since each recurrent depth can be read out independently, a single LoopLM exposes a broader output space across inference depths, raising an important question: whether safety is preserved throughout recurrent computation. Prior evaluations suggest that deeper recurrence can improve safety on harmful queries, but robustness under jailbreak attacks remains unclear. We therefore conduct a comprehensive safety evaluation of LoopLMs under jailbreak attacks targeting different recurrent depths. We find that attack success can increase at deeper inference depths, the same query can elicit different safety behaviors across depths, and attacks constructed against one depth can transfer to others. Moreover, SFT and preference alignment do not eliminate these safety gaps, motivating an alignment method designed for LoopLMs.
More
Looped Language Models (LoopLMs) provide a parameter efficient approach to scaling model capabilities through repeated use of shared parameters across recurrent steps. Since each recurrent depth can be read out independently, a single LoopLM exposes a broader output space across inference depths, raising an important question: whether safety is preserved throughout recurrent computation. Prior evaluations suggest that deeper recurrence can improve safety on harmful queries, but robustness under jailbreak attacks remains unclear. We therefore conduct a comprehensive safety evaluation of LoopLMs under jailbreak attacks targeting different recurrent depths. We find that attack success can increase at deeper inference depths, the same query can elicit different safety behaviors across depths, and attacks constructed against one depth can transfer to others. Moreover, SFT and preference alignment do not eliminate these safety gaps, motivating an alignment method designed for LoopLMs.
Read paper →
arXiv
Oct 7, 2026
Adam Pardyl, Siddhartha Gairola, Sukrut Rao, Adam Wróbel, Bartosz Zieliński, Bernt Schiele, Dawid Rymarczyk
Concept-based vision models represent images through an intermediate layer of human-inspectable concepts, so what a model relies on can be traced to those concepts. However, those models are often limited to fixed categories or depend on language to define their concepts. We introduce DisParQ (Discrete Parts with Quantized attributes), a method that learns spatially grounded, discrete concept representations from a powerful frozen vision-only self-supervised backbone. It requires no class labels and no language supervision. Each image patch is assigned to exactly one concept from a learnable prototype dictionary, and only a sparse subset of concepts may activate per image. To capture how each concept varies across images (e.g., the type of a "wheel"), we learn continuous residuals alongside the concepts and then quantize them into discrete attributes.
More
Concept-based vision models represent images through an intermediate layer of human-inspectable concepts, so what a model relies on can be traced to those concepts. However, those models are often limited to fixed categories or depend on language to define their concepts. We introduce DisParQ (Discrete Parts with Quantized attributes), a method that learns spatially grounded, discrete concept representations from a powerful frozen vision-only self-supervised backbone. It requires no class labels and no language supervision. Each image patch is assigned to exactly one concept from a learnable prototype dictionary, and only a sparse subset of concepts may activate per image. To capture how each concept varies across images (e.g., the type of a "wheel"), we learn continuous residuals alongside the concepts and then quantize them into discrete attributes.
Read paper →
LessWrong
Oct 5, 2026
Edward James Young
Research from LessWrong titled '“Alignment Engineering” vs. “Misalignment Science”'. Published 2026-10-05. Authors: Edward James Young.
More
Research from LessWrong titled '“Alignment Engineering” vs. “Misalignment Science”'. Published 2026-10-05. Authors: Edward James Young.
Read paper →
RAND
Oct 5, 2026
Expert Insights Oct 5, 2026 This perspective examines how federally funded research and development centers can close gaps across three domains: data and infrastructure, algorithmic intelligence, and autonomous experimentation.
More
Expert Insights Oct 5, 2026 This perspective examines how federally funded research and development centers can close gaps across three domains: data and infrastructure, algorithmic intelligence, and autonomous experimentation.
Read paper →
RAND
Oct 5, 2026
Research Oct 5, 2026 The U.S.-China tech rivalry hinges not only on innovation but also on achieving widespread technological diffusion in the economy. This report examines both countries’ capacity to convert technological progress into sustained productivity gains.
More
Research Oct 5, 2026 The U.S.-China tech rivalry hinges not only on innovation but also on achieving widespread technological diffusion in the economy. This report examines both countries’ capacity to convert technological progress into sustained productivity gains.
Read paper →
RAND
Oct 5, 2026
Research Oct 5, 2026 The authors applied RAND’s U.S. Code Analysis Tool (CAT) to the One Big Beautiful Bill Act as a proof of concept to show how CAT can support rapid statutory triage by identifying which statutory changes warrant closer review.
More
Research Oct 5, 2026 The authors applied RAND’s U.S. Code Analysis Tool (CAT) to the One Big Beautiful Bill Act as a proof of concept to show how CAT can support rapid statutory triage by identifying which statutory changes warrant closer review.
Read paper →
RAND
Oct 5, 2026
Research Oct 5, 2026 The authors synthesize observations from tabletop games on artificial intelligence conducted in Europe or with European scenarios and identify the key issues and capabilities that players deemed necessary to navigate AI’s potential adverse effects.
More
Research Oct 5, 2026 The authors synthesize observations from tabletop games on artificial intelligence conducted in Europe or with European scenarios and identify the key issues and capabilities that players deemed necessary to navigate AI’s potential adverse effects.
Read paper →