Robin Mills
Research Subjects
Research Interests
Robin Mills has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Aidan Philip Millar
Mubadala Investment Company, Abu Dhabi, United Arab Emirates
2 shared publications - Anup Roy
Inception42, Abu Dhabi, United Arab Emirates
2 shared publications - Larry Murray
Inception42, Abu Dhabi, United Arab Emirates
2 shared publications - Rishabh Upadhyay
University of Milano-Bicocca
2 shared publications - Ana Carolina Peçanha Antonio
Hospital de Clinicas de Porto Alegre, Porto Alegre, Brazil
1 shared publication - Animesh Rameshbhai Panara
Inception AI, Abu Dhabi, United Arab Emirates
1 shared publication
Latest Publications
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 7726-7731
PyRAMIND: Auditable Deterministic Pyramid Memory for Long-Running LLM Agents
Long-running LLM agents are increasingly considered for regulated, high-consequence domains such as finance, compliance, and audit, where memory must be accurate, traceable, replayable, and inspectable after a decision. Existing agent-memory systems often invoke an LLM during maintenance operations such as fact extraction, consolidation, updating, summarization, and forgetting. Although effective, this design makes it harder to reconstruct why a memory existed, changed, or was removed at a particular point in time. We present PyRAMIND (Pyramidal Retention with Arithmetic Memory for Inference-time Determinism), an auditable memory architecture in which write consolidation, reinforcement, forgetting, tier promotion, and retrieval scoring are closed-form arithmetic over cached embeddings. Only optional entity tagging at ingest and answer generation at query time call an external LLM. Given fixed embeddings, cached boundary-model outputs, parameters, and software, each maintenance transition is replayable from ordered inputs. PyRAMIND stores memories in a four-tier Pyramid (raw →facts →themes→worldview) and ranks them with a max-ensemble retention score modulated by local semantic density. Under a matched harness, PyRAMIND achieves competitive accuracy on long-horizon memory tasks (0.93 on LongMemEval-S) despite eliminating LLM calls from the memory-maintenance path, while reducing memory-lifecycle token expenditure to 1.4k LLM tokens per turn versus 3.1k for reproduced Mem0 OSS and 4.2k for Zep v3. Performance is weaker on temporal and multi-hop reasoning tasks (0.62 on LoCoMo), highlighting the continued value of richer relational memory structures for compositional reasoning. These results suggest that strong long-term memory performance is possible without LLM-mediated maintenance, and motivate future replayable memory systems that add temporal and relational structure while preserving auditable deterministic maintenance.
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 7726-7731
Open AccessPyRAMIND: Auditable Deterministic Pyramid Memory for Long-Running LLM Agents
Long-running LLM agents are increasingly considered for regulated, high-consequence domains such as finance, compliance, and audit, where memory must be accurate, traceable, replayable, and inspectable after a decision. Existing agent-memory systems often invoke an LLM during maintenance operations such as fact extraction, consolidation, updating, summarization, and forgetting. Although effective, this design makes it harder to reconstruct why a memory existed, changed, or was removed at a particular point in time. We present PyRAMIND (Pyramidal Retention with Arithmetic Memory for Inference-time Determinism), an auditable memory architecture in which write consolidation, reinforcement, forgetting, tier promotion, and retrieval scoring are closed-form arithmetic over cached embeddings. Only optional entity tagging at ingest and answer generation at query time call an external LLM. Given fixed embeddings, cached boundary-model outputs, parameters, and software, each maintenance transition is replayable from ordered inputs. PyRAMIND stores memories in a four-tier Pyramid (raw →facts →themes→worldview) and ranks them with a max-ensemble retention score modulated by local semantic density. Under a matched harness, PyRAMIND achieves competitive accuracy on long-horizon memory tasks (0.93 on LongMemEval-S) despite eliminating LLM calls from the memory-maintenance path, while reducing memory-lifecycle token expenditure to 1.4k LLM tokens per turn versus 3.1k for reproduced Mem0 OSS and 4.2k for Zep v3. Performance is weaker on temporal and multi-hop reasoning tasks (0.62 on LoCoMo), highlighting the continued value of richer relational memory structures for compositional reasoning. These results suggest that strong long-term memory performance is possible without LLM-mediated maintenance, and motivate future replayable memory systems that add temporal and relational structure while preserving auditable deterministic maintenance.
- Open Access
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 4087-4092
BANANA: Bounded Adaptive Navigation Architecture for Nested Archives
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 4087-4092
Open AccessBANANA: Bounded Adaptive Navigation Architecture for Nested Archives
- Open Access
Published benefits of ivermectin use in Itajaí, Brazil for COVID-19 infection, hospitalisation, and mortality are entirely explained by statistical artefacts
Background Two recent publications by Kerr et al. (Cureus 14(1):e21272; Cureus 14(8): e28624) reported dramatic effects of prophylactic ivermectin use for both prevention of COVID-19 and reduction of COVID-19-related hospitalisation and mortality, including a dose-dependent effect of ivermectin prophylaxis. These papers have gained an unusually large public influence: they were incorporated into debates around COVID-19 policies and may have contributed to decreased trust in vaccine efficacy and public health authorities more broadly. Both studies were based on retrospective observational analysis of city-wide registry data from the city of Itajaí, Brazil from July-December 2020. Methods Starting with initially identified sources of error, we conducted a revised statistical analysis of available data, including data made available with the original papers and public data from the Brazil Ministry of Health. We identified additional uncorrected sources of bias and errors from the original analysis, including incorrect subject exclusion and missing subjects, an enrolment time bias, and multiple sources of immortal time bias. In models assuming no actual effect from ivermectin use, we conducted Monte Carlo simulations to estimate the contribution of these biases to any observed effect. Results Untreated statistical artefacts and methodological errors alone lead to dramatic apparent risk reduction associated with ivermectin use in both studies. The magnitude of apparent risk reduction from these artefacts is comparable to the results reported by the studies themselves, including apparent protection from infection, hospitalisation, and death, and including the reported apparent dose-response relationship. Conclusions The inference of ivermectin efficacy reported in both papers is unsupported, as the observed effects are entirely explained by untreated statistical artefacts and methodological errors. Our re-analysis calls for caution in interpreting highly publicised observational studies and highlights the importance of common sources of bias in clinical research.
Open AccessPublished benefits of ivermectin use in Itajaí, Brazil for COVID-19 infection, hospitalisation, and mortality are entirely explained by statistical artefacts
Background Two recent publications by Kerr et al. (Cureus 14(1):e21272; Cureus 14(8): e28624) reported dramatic effects of prophylactic ivermectin use for both prevention of COVID-19 and reduction of COVID-19-related hospitalisation and mortality, including a dose-dependent effect of ivermectin prophylaxis. These papers have gained an unusually large public influence: they were incorporated into debates around COVID-19 policies and may have contributed to decreased trust in vaccine efficacy and public health authorities more broadly. Both studies were based on retrospective observational analysis of city-wide registry data from the city of Itajaí, Brazil from July-December 2020. Methods Starting with initially identified sources of error, we conducted a revised statistical analysis of available data, including data made available with the original papers and public data from the Brazil Ministry of Health. We identified additional uncorrected sources of bias and errors from the original analysis, including incorrect subject exclusion and missing subjects, an enrolment time bias, and multiple sources of immortal time bias. In models assuming no actual effect from ivermectin use, we conducted Monte Carlo simulations to estimate the contribution of these biases to any observed effect. Results Untreated statistical artefacts and methodological errors alone lead to dramatic apparent risk reduction associated with ivermectin use in both studies. The magnitude of apparent risk reduction from these artefacts is comparable to the results reported by the studies themselves, including apparent protection from infection, hospitalisation, and death, and including the reported apparent dose-response relationship. Conclusions The inference of ivermectin efficacy reported in both papers is unsupported, as the observed effects are entirely explained by untreated statistical artefacts and methodological errors. Our re-analysis calls for caution in interpreting highly publicised observational studies and highlights the importance of common sources of bias in clinical research.
- Open Access
- 20 March 2019
- Published by Wiley in Journal of Avian Biology
- Vol. 50 (3)
Sexual size dimorphism, prey morphology and catch success in relation to flight mechanics in the peregrine falcon: a simulation study
In common with many other raptors, female peregrine falcons Falco peregrinus are about 50% heavier than males. Their sexual dimorphism is thought to allow breeding pairs to exploit a wider range of prey through a division of labor: the male being able to catch more maneuverable prey species; the female capable of carrying larger ones. Given the difficulty of assessing the catch success and load carrying capacity of both sexes of falcon in the field, we here adopt a novel approach to test the division‐of‐labor theory by using a detailed physics‐based flight simulator of birds. We study attacks by male and female peregrines on prey species ranging from small passerines to large ducks, testing how catch success relates to the flight performance of predator and prey. Males prove to be better than females at catching highly maneuverable prey in level flight, but the catch success of both sexes improves and becomes more similar when diving, because of the higher aerodynamic forces that are available to both sexes for maneuvering in high‐speed flight. The higher maximum roll acceleration of the male peregrine explains its edge over the female in catching maneuverable prey in level flight. Overall, catch success is more strongly influenced by the differences in maneuverability that exist between different species of prey than between the different sexes of falcon. On the other hand, the female can carry up to 50% greater loads than the male. More generally, our detailed simulation approach highlights the importance of several previously overlooked features of attack and escape. In particular, we find that it is not the prey's instantaneous maximum centripetal acceleration but the prey's ability to sustain a high centripetal acceleration for an extended period of time that is the primary driver of the variation in catch success across species.
- 20 March 2019
- Published by Wiley in Journal of Avian Biology
- Vol. 50 (3)
Open AccessSexual size dimorphism, prey morphology and catch success in relation to flight mechanics in the peregrine falcon: a simulation study
In common with many other raptors, female peregrine falcons Falco peregrinus are about 50% heavier than males. Their sexual dimorphism is thought to allow breeding pairs to exploit a wider range of prey through a division of labor: the male being able to catch more maneuverable prey species; the female capable of carrying larger ones. Given the difficulty of assessing the catch success and load carrying capacity of both sexes of falcon in the field, we here adopt a novel approach to test the division‐of‐labor theory by using a detailed physics‐based flight simulator of birds. We study attacks by male and female peregrines on prey species ranging from small passerines to large ducks, testing how catch success relates to the flight performance of predator and prey. Males prove to be better than females at catching highly maneuverable prey in level flight, but the catch success of both sexes improves and becomes more similar when diving, because of the higher aerodynamic forces that are available to both sexes for maneuvering in high‐speed flight. The higher maximum roll acceleration of the male peregrine explains its edge over the female in catching maneuverable prey in level flight. Overall, catch success is more strongly influenced by the differences in maneuverability that exist between different species of prey than between the different sexes of falcon. On the other hand, the female can carry up to 50% greater loads than the male. More generally, our detailed simulation approach highlights the importance of several previously overlooked features of attack and escape. In particular, we find that it is not the prey's instantaneous maximum centripetal acceleration but the prey's ability to sustain a high centripetal acceleration for an extended period of time that is the primary driver of the variation in catch success across species.
- Open Access
- 12 April 2018
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 14 (4) , e1006044
Physics-based simulations of aerial attacks by peregrine falcons reveal that stooping at high speed maximizes catch success against agile prey
The peregrine falcon Falco peregrinus is renowned for attacking its prey from high altitude in a fast controlled dive called a stoop. Many other raptors employ a similar mode of attack, but the functional benefits of stooping remain obscure. Here we investigate whether, when, and why stooping promotes catch success, using a three-dimensional, agent-based modeling approach to simulate attacks of falcons on aerial prey. We simulate avian flapping and gliding flight using an analytical quasi-steady model of the aerodynamic forces and moments, parametrized by empirical measurements of flight morphology. The model-birds’ flight control inputs are commanded by their guidance system, comprising a phenomenological model of its vision, guidance, and control. To intercept its prey, model-falcons use the same guidance law as missiles (pure proportional navigation); this assumption is corroborated by empirical data on peregrine falcons hunting lures. We parametrically vary the falcon’s starting position relative to its prey, together with the feedback gain of its guidance loop, under differing assumptions regarding its errors and delay in vision and control, and for three different patterns of prey motion. We find that, when the prey maneuvers erratically, high-altitude stoops increase catch success compared to low-altitude attacks, but only if the falcon’s guidance law is appropriately tuned, and only given a high degree of precision in vision and control. Remarkably, the optimal tuning of the guidance law in our simulations coincides closely with what has been observed empirically in peregrines. High-altitude stoops are shown to be beneficial because their high airspeed enables production of higher aerodynamic forces for maneuvering, and facilitates higher roll agility as the wings are tucked, each of which is essential to catching maneuvering prey at realistic response delays. Peregrine falcons are famed for their high-speed, high-altitude stoops. Hunting prey at perhaps the highest speed of any animal places a stooping falcon under extraordinary physical, physiological, and cognitive demands, yet it remains unknown how this behavioural strategy promotes catch success. Because the behavioral aspects of stooping are intimately related to its biomechanical constraints, we address this question through an embodied cognition approach. We model the falcon’s cognition using guidance laws inspired by theory and experiment, and embody this in a physics-based simulation of predator and prey flight. Stooping maximizes catch success against agile prey by minimizing roll inertia and maximizing the aerodynamic forces available for maneuvering, but requires a tightly tuned guidance law, and exquisitely precise vision and control.
- 12 April 2018
- Published by Public Library of Science (PLoS) in PLOS Computational Biology
- Vol. 14 (4) , e1006044
Open AccessPhysics-based simulations of aerial attacks by peregrine falcons reveal that stooping at high speed maximizes catch success against agile prey
The peregrine falcon Falco peregrinus is renowned for attacking its prey from high altitude in a fast controlled dive called a stoop. Many other raptors employ a similar mode of attack, but the functional benefits of stooping remain obscure. Here we investigate whether, when, and why stooping promotes catch success, using a three-dimensional, agent-based modeling approach to simulate attacks of falcons on aerial prey. We simulate avian flapping and gliding flight using an analytical quasi-steady model of the aerodynamic forces and moments, parametrized by empirical measurements of flight morphology. The model-birds’ flight control inputs are commanded by their guidance system, comprising a phenomenological model of its vision, guidance, and control. To intercept its prey, model-falcons use the same guidance law as missiles (pure proportional navigation); this assumption is corroborated by empirical data on peregrine falcons hunting lures. We parametrically vary the falcon’s starting position relative to its prey, together with the feedback gain of its guidance loop, under differing assumptions regarding its errors and delay in vision and control, and for three different patterns of prey motion. We find that, when the prey maneuvers erratically, high-altitude stoops increase catch success compared to low-altitude attacks, but only if the falcon’s guidance law is appropriately tuned, and only given a high degree of precision in vision and control. Remarkably, the optimal tuning of the guidance law in our simulations coincides closely with what has been observed empirically in peregrines. High-altitude stoops are shown to be beneficial because their high airspeed enables production of higher aerodynamic forces for maneuvering, and facilitates higher roll agility as the wings are tucked, each of which is essential to catching maneuvering prey at realistic response delays. Peregrine falcons are famed for their high-speed, high-altitude stoops. Hunting prey at perhaps the highest speed of any animal places a stooping falcon under extraordinary physical, physiological, and cognitive demands, yet it remains unknown how this behavioural strategy promotes catch success. Because the behavioral aspects of stooping are intimately related to its biomechanical constraints, we address this question through an embodied cognition approach. We model the falcon’s cognition using guidance laws inspired by theory and experiment, and embody this in a physics-based simulation of predator and prey flight. Stooping maximizes catch success against agile prey by minimizing roll inertia and maximizing the aerodynamic forces available for maneuvering, but requires a tightly tuned guidance law, and exquisitely precise vision and control.