ATLAS steers AI agents’ hidden states to reduce coordination drift

A team of specialized AI agents needs more than individually fluent responses. Its planner must follow the current objective, its memory writer must preserve updates, and its executor must use the latest information without taking over another agent’s responsibilities. Over a long interaction, those boundaries can blur.

In an arXiv preprint posted on November 3, 2026, Vishal Pramanik, Maisha Maliha, Susmit Jha and Sumit Kumar Jha introduce ATLAS, short for Algebraic Task–role Locking with Adaptive Spectral Memory. The supplied manuscript states that it was accepted at EMNLP 2026. Its central hypothesis is that task drift, role drift and stale-state use share a representational pattern inside language models.

Looking beneath the conversation

The researchers first studied 14,832 annotated turns from 200 ten-agent Qwen3-8B conversations. Three model evaluators classified departures from the active task, assigned authority or current shared state. Turns labeled as drifted showed lower alignment between internal model representations and subspaces constructed from descriptions of those three constraints.

The failure types were correlated, with pairwise correlations between 0.72 and 0.78. This supports an association, not by itself proof of a common cause. The paper adds temporal observations and interventions that remove selected representational directions to strengthen its case.

There is also a measurement caveat: the alignment diagnostic uses the same anchor construction that ATLAS steers toward. The authors acknowledge this dependence and provide random-subspace controls. A separate, blinded human audit of 200 turns agreed with the model-majority drift labels on 89.5% of turns, supporting the labeling procedure without eliminating disagreement.

A correction before each token

ATLAS leaves the language model’s weights unchanged. It processes task descriptions, role definitions and shared-state summaries through the frozen model, then extracts compact subspaces from the resulting hidden states at a selected intermediate layer.

These anchors are combined into a positive-semidefinite kernel: a mathematical object that assigns greater influence to directions supported by multiple anchors. Instead of forcing task, role and state into separate, non-overlapping compartments, the method preserves their overlap.

During generation, ATLAS modestly amplifies a token’s hidden-state components along the kernel’s weighted directions before passing the representation to downstream layers. A gain cap limits the strongest per-direction correction. Low-rank computation avoids constructing a full hidden-dimension-sized matrix.

The state anchor evolves after each turn. Existing state directions lose importance, while sufficiently novel residual information is added. Rank limits prevent indefinite growth. In heterogeneous teams, every agent constructs its anchors within its own model; the method does not require a common hidden coordinate system across model families.

What the benchmarks show

The main experiments cover long-conversation memory in EverMemBench and interactive state maintenance in AMemGym. They compare a ten-agent homogeneous team, a twelve-agent homogeneous team and a twelve-agent team spanning six model families. Reported results average three seeds under shared episode-level context budgets.

On EverMemBench, ATLAS exceeds the adapted MemoryArena-Agent baseline by 4.2–4.8 percentage points across those settings. In the twelve-agent Qwen3-8B setting, its macro-average is 54.1%, compared with 49.3% for MemoryArena-Agent.

On AMemGym in that setting, overall accuracy rises from 55.4% to 60.1%. Write failures fall from 15.6% to 11.4%, and read failures from 13.2% to 10.6%. Utilization failures, however, increase from 16.2% to 17.9%: remembering and retrieving more information does not ensure that the executor uses it correctly.

A larger-backbone experiment reports an EverMemBench macro-average of 61.0% with Qwen2.5-72B-Instruct, versus 56.2% for MemoryArena-Agent. Supplementary tool-use, least-privilege and simulated robotic stress tests also report gains, but remain controlled evaluations rather than evidence of unrestricted deployment reliability.

What the gains establish—and what they do not

The complete system matters. Removing spectral steering while retaining ATLAS’s other mechanisms lowers the Qwen3-8B memory score from 54.1% to 52.1%. That two-point difference measures steering’s marginal contribution within this pipeline, not the entire improvement over competing systems.

Comparisons with activation-steering baselines also have a coverage difference: ATLAS steers every active agent, while ActAdd and SEA intervene on three designated roles. These experiments therefore compare configurations, not steering rules under matched coverage.

ATLAS needs access to intermediate hidden states, making it unsuitable as implemented for black-box APIs. Its anchors can also encode poor specifications: contradictory descriptions and swapped roles substantially degrade performance. The practical result is a promising internal coordination control—not a replacement for factual verification, enforced permissions, external safety checks or human oversight.