Turn: A Multidisciplinary Reorientation Unit

• Turn is defined as a fundamental unit representing interaction, reorientation, or process transition across diverse fields such as dialogue systems, RL agents, and physical systems.
• In dialogue and agentic reinforcement learning, turn-level modeling encapsulates complete agent–environment interactions, enabling optimized credit assignment and improved decision-making.
• In physical and geometric contexts, turn quantifies directional changes and structural reorientations, serving as a critical index in accelerator diagnostics, fluid mechanics, and quantum control.

Searching arXiv for the cited works on “turn” across dialogue systems, agent RL, fluid mechanics, and geometric/physical uses. Across the cited literature, “turn” denotes a family of formally distinct but structurally related concepts: an interaction round in dialogue and agentic reinforcement learning, a turn-taking state in spoken systems, a revolution index in accelerator diagnostics, a local turning axis in fluid geometry, a topological or morphogenetic reorientation in biological systems, a controlled directional change in wave and active-matter transport, and a geometric or combinatorial object in quantum control and lattice geometry. In each case, the term identifies a unit at which progression, reorientation, or credit assignment becomes technically meaningful, and the associated mathematics is chosen to make that unit observable, optimizable, or classifiable (Cooper et al., 18 May 2026, Tan et al., 26 Jun 2026, Simon et al., 2011, Simon et al., 2011).

1. Turn as an interaction unit in multi-turn language-model agents

In several recent RL formulations for LLM agents, a turn is defined as one complete agent–environment interaction step rather than as an arbitrary token span. ATOD formalizes a trajectory as

$\tau = (x, a_1, o_1, \ldots, a_K, o_K, a_{K+1}),$

where $a_k$ is the model response at step $k$ and $o_k$ is the returned observation; in that setting, a turn is “the natural interaction unit of agent behavior,” namely one model response associated with a decision step in the environment or tool-using process (Tan et al., 26 Jun 2026). Turn structure is then made algorithmically explicit through Turn-level Disagreement-Uncertainty Reweighting, with

$A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$

so that token-level distillation is modulated by the turn containing the token (Tan et al., 26 Jun 2026).

A closely related move appears in turn-level RL for agentic reasoning. “Turn-PPO” reformulates the MDP so that the state at turn $n$ is the accumulated history plus the current environment query,

$$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$

and the action is the full response

$a_n := R_n,$

thereby replacing token-level credit assignment by turn-level advantage estimation (Li et al., 18 Dec 2025). The paper’s argument is that multi-turn interaction alternates between model-generated tokens and externally injected environment tokens, so turn-level abstraction better matches the agent–environment loop and makes $\gamma$ and $\lambda$ meaningfully tunable in GAE (Li et al., 18 Dec 2025).

Other works make the same claim through different optimization designs. A$a_k$0TGPO treats each model-produced segment $a_k$1 as the operative interaction unit, computes Information Gain per turn,

$a_k$2

normalizes it within same-depth turn groups, and uses a turn-level importance ratio in a PPO-style objective (Chen et al., 7 May 2026). GTPO, designed for tool-integrated mathematical reasoning, defines a TIR trajectory as

$a_k$3

with $a_k$4 a reasoning-and-tool-use turn and $a_k$5 the tool feedback, then assigns turn-level rewards $a_k$6, discounted turn returns

$a_k$7

and turn-level group-normalized advantages $a_k$8 (Ding et al., 18 Nov 2025). “Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Reward Design” makes the same unit explicit at the MDP level: $a_k$9 with turn-level rewards $k$0 and rollouts

$k$1

where a turn is “one complete agent-environment interaction, such as a tool call and its result” (Wei et al., 17 May 2025).

A plausible implication is that, across agent RL papers, “turn” functions as a macro-action abstraction: coarser than a token, finer than a whole trajectory, and especially useful when external observations or tool outputs intervene between model-generated segments. The cited works do not claim a universal formulation for all agentic systems, but they consistently treat the turn as the scale at which semantically coherent decisions, delayed feedback, and process credit become tractable (Tan et al., 26 Jun 2026, Chen et al., 7 May 2026, Li et al., 18 Dec 2025, Ding et al., 18 Nov 2025, Wei et al., 17 May 2025).

2. Turn structure in dialogue systems, turn-taking, safety, and backdoors

In dialogue research, “turn” retains its conversational meaning—one user-assistant exchange—but is operationalized in sharply different ways depending on the task. “Easy Turn” defines turn-taking for full-duplex spoken dialogue systems as a four-state prediction problem over user speech: complete, incomplete, backchannel, and wait (Li et al., 28 Sep 2025). The model takes speech plus a natural-language prompt and generates an ASR transcription followed by a turn-state label, using a Whisper-Medium encoder, a CNN/Transformer adaptor, and Qwen2.5-0.5B-Instruct in an ASR+Turn-Detection paradigm (Li et al., 28 Sep 2025). In that usage, a turn is neither merely silence endpointing nor an arbitrary token chunk, but an interactionally meaningful control state for deciding whether the system should respond, keep listening, ignore feedback, or halt (Li et al., 28 Sep 2025).

In safety work, the turn becomes a temporal substrate for latent malicious intent. “Speak Out of Turn” studies multi-turn dialogue as a vulnerability class in which a harmful objective is decomposed into benign-looking sub-queries spread across turns, so that no individual turn necessarily appears overtly unsafe even though the whole dialogue becomes harmful (Zhou et al., 2024). The attack methodology is “Malicious Query Decomposition,” and the paper reports sharply higher harmfulness for multi-turn decompositions than for direct malicious queries or one-shot concatenations of the same sub-queries (Zhou et al., 2024). “One Turn Too Late” reframes the defense problem at the same granularity: the defender observes

$k$2

defines a sufficiency operator $k$3, and seeks the earliest harmful closure turn

$k$4

that is, the earliest turn at which delivering the candidate response would make the interaction harm-enabling (Shen et al., 7 May 2026). TurnGate then learns a response-aware Pass/Block policy over turns using supervised warm-starting and a PPO-style offline RL objective (Shen et al., 7 May 2026).

A different security use appears in backdoor attacks. “Turn-Based Structural Triggers” defines a Turn-based Structural Trigger $k$5, often the even turns $k$6, so that the backdoored model behaves as

$k$7

and the trigger is the dialogue turn index rather than a user-visible lexical pattern (Lu et al., 20 Jan 2026). The paper reports average ASR $k$8 across four open-source LLMs with minimal utility degradation, and sustained effectiveness under five representative defenses with average ASR $k$9 (Lu et al., 20 Jan 2026). This makes turn structure itself an attack surface.

The common theme is that dialogue turns are not only sequencing devices. They can be the unit of system control, the unit of distributed intent, the unit of earliest harmful sufficiency, or the latent index used for structural triggering. This suggests that turn-aware modeling in dialogue safety cannot be reduced to prompt-local content moderation alone (Li et al., 28 Sep 2025, Zhou et al., 2024, Shen et al., 7 May 2026, Lu et al., 20 Jan 2026).

3. Turn as iteration, refinement, and serial structure

A different line of work treats turns not as dialogue exchanges between distinct agents, but as ordered iterations of self-refinement. “Another Turn, Better Output?” runs controlled 12-turn conversations in ideation, code, and math, with Turn 1 producing the initial answer and Turns 2–12 conditioning only on the model’s immediately previous response through the template “The following is a previous response: …” plus an improvement instruction (Javaji et al., 8 Sep 2025). The paper tracks turn-level behavior using Drift from Origin,

$o_k$0

turn-to-turn Volatility,

$o_k$1

Lexical Novelty, and Growth Factor

$o_k$2

with $o_k$3 instantiated as word count for ideation and math and lines of code for coding (Javaji et al., 8 Sep 2025). The main result is strongly domain-dependent: gains arrive early in ideas and code, while in math late turns matter when guided by elaboration (Javaji et al., 8 Sep 2025).

A related but adversarial use of serial structure appears in “One-Shot is Enough,” which compresses multi-turn jailbreak conversations into single-turn prompts via Hyphenize, Numberize, and Pythonize (Ha et al., 6 Mar 2025). The paper formalizes the original multi-turn dialogue as

$o_k$4

and then replaces it with a single serialized prompt

$o_k$5

On the Multi-turn Human Jailbreak dataset, the M2S methods achieve ASR from $o_k$6 to $o_k$7 across the tested LLMs, and the M2S ensemble outperforms the original multi-turn attacks on all four base models (Ha et al., 6 Mar 2025). The paper interprets this as evidence that much of the adversarial potency of a turn sequence can survive serialization into one prompt.

Taken together, these works treat turn order as a measurable or exploitable trajectory in its own right. In iterative refinement, turn metrics indicate when to steer, stop, or switch strategy (Javaji et al., 8 Sep 2025). In M2S jailbreak compression, the turn sequence is treated as a structure that can be preserved even when interactive exchange is removed (Ha et al., 6 Mar 2025). A plausible implication is that “turn” often functions less as a strictly interpersonal notion than as a general serial scaffold for staged instruction.

4. Turn as a temporal index in beam and measurement systems

In accelerator diagnostics, “turn” denotes the revolution index of a circulating beam. “Tune Evaluation From Phased BPM Turn-By-Turn Data” defines turn-by-turn BPM data as one beam position sample per revolution turn, indexed by BPM number $o_k$8 and turn number $o_k$9, and studies tune extraction in a fast-ramping synchrotron where the tune can change significantly over only a few tens of turns (Alexahin et al., 2012). Because a long FFT is incompatible with rapidly varying tune, the paper uses a Continuous Fourier Transform over turn number,

$A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$0

and then improves SNR by phasing and summing many BPMs on each machine turn: $A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$1 The coherent betatron component is amplified by $A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$2, whereas alien modes and random noise scale only as $A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$3, yielding an SNR gain of $A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$4 (Alexahin et al., 2012).

In this literature, a turn is the fundamental time index of oscillation and the synchronization unit on which spatial BPM phasing is performed. It is neither a geometric bend nor an interaction round, but the discrete recurrence of the beam after one revolution. The paper’s solution depends on exploiting both dimensions of the data simultaneously: across BPMs on a given turn, and across turns after a kick (Alexahin et al., 2012).

This use of “turn” is temporally elementary and mechanically precise. It suggests a broader pattern across fields: the term often labels the minimal recurrence interval at which a complex process becomes analyzable. In synchrotron tune measurement, that interval is the machine revolution (Alexahin et al., 2012).

5. Turn as local geometric orientation and large-scale morphological transformation

In fluid mechanics, “turn” can denote the local orientation of streamline bending. “Mapping the Turn” introduces an Eulerian binormal-axis diagnostic for three-dimensional recirculating flows, defined from the velocity field $A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$5 and its convective acceleration

$A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$6

through

$A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$7

Using the Frenet–Serret decomposition, the paper shows that

$A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$8

so wherever speed and curvature are nonzero, the normalized direction recovers the streamline binormal, i.e. the local axis normal to the osculating plane of the streamline (Cooper et al., 18 May 2026). The method then projects $A^{OPD}_t=\Delta\log p_t\,w_{k(t)},$9 onto streamwise, wall-normal, and spanwise axes using barycentric RGB weights

$n$0

producing a spatially resolved field of turning-axis orientation in flows such as Hill’s spherical vortex and a pressure-gradient-induced three-dimensional separation bubble (Cooper et al., 18 May 2026).

In developmental mechanics, turning is literal topological inversion. “Dynamics of a Volvox Embryo Turning Itself Inside Out” studies inversion in Volvox globator through SPIM imaging and elastic-shell theory (Höhn et al., 2014). The embryo begins as a hollow monolayer whose flagellated cell surfaces face inward, so inversion is required to complete embryogenesis (Höhn et al., 2014). The shell model uses stretches

$n$1

intrinsic stretches $n$2, intrinsic curvatures $n$3, and elastic energy

$n$4

with posterior contraction and a localized equatorial band of negative preferred curvature required to reproduce the observed “mushroom” morphologies (Höhn et al., 2014). The same paper also models a snap-through-like speed-up resisted by surrounding fluid and derives a reduced dynamics

$n$5

consistent with an experimentally observed abrupt increase in inversion speed (Höhn et al., 2014).

Both works quantify turning as geometry rather than as trajectory count: in one case, the local axis about which a streamline bends; in the other, the global inside-out transformation of a deforming shell. This suggests that “turn” in continuum systems often marks a change in orientation class that is more naturally described by curvature, osculating planes, intrinsic geometry, or topological inversion than by discrete state labels (Cooper et al., 18 May 2026, Höhn et al., 2014).

6. Turn as directed reorientation, routing, and geometric representation

In wave and active-matter systems, turn denotes controlled reorientation under anisotropic constraints. “Zero-field spin wave turns” studies the routing of dipolar spin waves through $n$6 bends in corrugated Permalloy waveguides with engineered in-plane magnetization landscapes (Klíma et al., 2023). Because the component of the wave vector parallel to an interface is preserved and dipolar spin waves have anisotropic dispersion, ordinary Damon–Eshbach-like bends oversteer and fail (Klíma et al., 2023). The paper instead chooses magnetization and interface orientations so that the group velocity follows the bend, and in the refined WG3 design selects the interface as a symmetry axis between incoming and refracted $n$7, thereby conserving $n$8 while rotating its direction (Klíma et al., 2023). Phase-resolved BLS then shows linear phase evolution through the turn, with projected phase slope $n$9, corresponding to $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$0 and $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$1 (Klíma et al., 2023).

“Anisotropic run-and-tumble-turn dynamics” uses “turn” in a stochastic transport sense. Orientation $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$2 evolves under

$$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$3

with a four-well angular potential whose minima at $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$4 make the elementary reorientation event a $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$5 tumble-turn (Loewe et al., 2023). Barrier crossing yields a Kramers-like turn rate $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$6, and the long-time dynamics collapse when normalized by

$$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$7

so that the coarse-grained process becomes a four-state directional switching model with hydrodynamic density equation

$$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$8

showing that diffusion emerges before isotropy (Loewe et al., 2023).

In quantum control, Hamilton’s turns are neither kinematic events nor stochastic reorientations, but geometric representatives of $$s_n := \left(\oplus_{n'<n}(Q_{n'}, R_{n'})\right) \oplus Q_n,$$9 elements. “Hamilton’s turns as visual tool-kit for designing of single-qubit unitary gates” associates each unitary

$a_n := R_n,$0

with a directed great-circle arc

$a_n := R_n,$1

on the unit sphere, where a Hamilton turn is the equivalence class of such arcs under rigid sliding along the same great circle (Simon et al., 2011). Multiplication of $a_n := R_n,$2 matrices becomes addition of turns: $a_n := R_n,$3 a non-Abelian analogue of the parallelogram law (Simon et al., 2011). The paper then uses this calculus to show that every single-qubit gate can be realized with two QWPs and one HWP (Simon et al., 2011).

A combinatorial version appears in “Reachability of turn sequences,” where a turn sequence

$a_n := R_n,$4

is realized as a simple rectilinear chain of integral segments starting at the origin and initially directed east (Evans et al., 2022). The endpoint set $a_n := R_n,$5 is characterized via the Stretching Lemma and Axis Lemma, with signed-axis reachability exactly equivalent to the presence of the corresponding hook type: $a_n := R_n,$6 iff a right hook exists, $a_n := R_n,$7 iff a left hook exists, $a_n := R_n,$8 iff an up hook exists, and $a_n := R_n,$9 iff a down hook exists (Evans et al., 2022). Distance bounds then depend on the excess number

$\gamma$0

and, in some $\gamma$1-axis cases, on maximal monotone prefixes or suffixes (Evans et al., 2022).

Across these otherwise unrelated fields, “turn” denotes a reorientation primitive constrained by structure: anisotropic dispersion in magnonics, four-well angular potentials in active matter, great-circle arc composition in $\gamma$2, and left/right combinatorics in lattice chains. This suggests that one recurring technical function of the term is to mark the smallest orientation-changing unit from which larger routing, control, or reachability theories can be built (Klíma et al., 2023, Loewe et al., 2023, Simon et al., 2011, Evans et al., 2022).